Final FLaReNet Deliverable Language Resources for the Future

Comments

Transcription

Final FLaReNet Deliverable Language Resources for the Future
The FLaReNet Book
ECP-2007-LANG-617001
FLaReNet
Final FLaReNet Deliverable
Language Resources for the Future – The
Future of Language Resources
The Strategic Language Resource Agenda
Deliverable number/name
D2.2b, D4.2, D4.3, D5.2, D6.2, D7.2, D7.3,
D8.2c
Dissemination level
Public
Delivery date
30 September 2011
Status
Final
Author(s)
Nicoletta Calzolari, Nuria Bel, Khalid
Choukri, Joseph Mariani, Monica Monachini,
Jan Odijk, Stelios Piperidis, Valeria Quochi,
Claudia Soria
eContentplus
This
project
is
funded
under
the
eContentplus
programme 1,
a multiannual Community programme to make digital content in Europe more accessible, usable and
exploitable.
1 OJ L 79, 24.3.2005, p. 1.
1
Table of Contents
The FLaReNet Book
TABLE OF CONTENTS ......................................................................................................................................................... 2
INTRODUCTION ................................................................................................................................................................... 4
THE CONTEXT............................................................................................................................................................................................4
WHAT ARE LANGUAGE RESOURCES ......................................................................................................................................................4
THE FLARENET MISSION, POLICY AND STRATEGY .............................................................................................................................5
THE FLARENET COMMUNITY ................................................................................................................................................................5
THE FLARENET DATA.............................................................................................................................................................................6
THE FLARENET RECOMMENDATIONS: A COLLECTIVE ENTERPRISE................................................................................................6
ORGANISATION OF THE BOOK ................................................................................................................................................................6
ACKNOWLEDGEMENTS ..................................................................................................................................................... 7
CHAPTER 1. THE FLARENET BLUEPRINT OF ACTIONS AND INFRASTRUCTURES ....................................... 8
HOW TO USE THE BLUEPRINT ................................................................................................................................................................9
1.1 RECOMMENDATIONS AT A GLANCE ................................................................................................................................................9
1.2
INFRASTRUCTURE OF LANGUAGE RESOURCES.................................................................................................................. 13
1.3
RESOURCE DOCUMENTATION .............................................................................................................................................. 14
1.4
RESOURCE DEVELOPMENT................................................................................................................................................... 17
1.5
RESOURCE INTEROPERABILITY ........................................................................................................................................... 20
1.6
RESOURCE COVERAGE, QUALITY, ADEQUACY ................................................................................................................... 22
1.6.1 RESOURCE QUANTITY ........................................................................................................................................................... 22
1.6.2 RESOURCE QUALITY .............................................................................................................................................................. 23
1.6.3 RESOURCE ADEQUACY .......................................................................................................................................................... 25
1.7
RESOURCE AVAILABILITY: SHARING AND DISTRIBUTION.............................................................................................. 26
1.8
RESOURCE SUSTAINABILITY ................................................................................................................................................ 28
1.9
RESOURCE RECOGNITION ..................................................................................................................................................... 29
1.10 INTERNATIONAL COOPERATION ......................................................................................................................................... 30
REFERENCES .......................................................................................................................................................................................... 31
CHAPTER 2 - IDENTIFYING MATURE AND SUSTAINABLE LANGUAGE RESOURCES ................................. 32
2.1
KEY SUSTAINABILITY FACTORS............................................................................................................................................ 34
2.2
KEY AREAS OF RESEARCH INTEREST ................................................................................................................................... 36
2.3
LR MATURITY AND SUSTAINABILITY IN MAJOR RESEARCH AREAS ................................................................................. 37
2.4
EVALUATION PACKAGES AND RESOURCES ......................................................................................................................... 39
2.5
MULTIPLE-APPLICATION RESOURCES AS A "MATURITY/SUSTAINABILITY" FACTOR .................................................. 41
2.6
MATURITY AS REFLECTED BY THE LRE MAP (2010) DATA ........................................................................................ 41
2.7
MATURITY IN TERMS OF OBJECTIVE EVALUATIONS .......................................................................................................... 43
2.8
FINAL RECOMMENDATIONS.................................................................................................................................................. 45
REFERENCES .......................................................................................................................................................................................... 46
CHAPTER 3 - STRATEGIES FOR LANGUAGE RESOURCE MIXING, SHARING, REUSING AND RECYCLING
................................................................................................................................................................................................ 47
3.1
REUSING RESOURCES............................................................................................................................................................. 47
3.2
INTERLINKING RESOURCES................................................................................................................................................... 48
3.3
REPURPOSING RESOURCES ................................................................................................................................................... 48
3.4
GENERAL CONSIDERATIONS ON REUSE............................................................................................................................... 49
REFERENCES .......................................................................................................................................................................................... 54
CHAPTER 4 - A STRATEGIC ACTION PLAN FOR AN INTEROPERABILITY FRAMEWORK ........................ 56
2
4.1
THE CURRENT STANDARD FRAMEWORK .......................................................................................................................... 56
4.2
BARRIERS AND MAJOR PROBLEMS....................................................................................................................................... 58
4.3
SCENARIOS FOR USING STANDARDS .................................................................................................................................... 59
4.4
RECOMMENDATIONS AND NEXT STEPS .............................................................................................................................. 61
4.5
TOWARDS AN INTEROPERABILITY FRAMEWORK ............................................................................................................. 65
REFERENCES .......................................................................................................................................................................................... 67
The FLaReNet Book
CHAPTER 5 - THE EVALUATION AND VALIDATION OF RESOURCES .............................................................. 70
5.1
5.2
RECOMMENDATIONS FOR VALIDATION.............................................................................................................................. 70
RECOMMENDATIONS FOR EVALUATION............................................................................................................................. 74
CHAPTER 6 - STRATEGIC DIRECTIONS FOR AUTOMATING LR DEVELOPMENT ........................................ 78
6.1
SURVEYING THE STATE OF THE ART.................................................................................................................................... 79
6.2
STRATEGIC DIRECTIONS AND RECOMMENDATIONS ........................................................................................................ 81
REFERENCES .......................................................................................................................................................................................... 84
CHAPTER 7 - TOWARDS AN LR ROADMAP .............................................................................................................. 85
7.1
URGENT LR REQUIREMENTS ............................................................................................................................................... 86
7.2
FINDINGS................................................................................................................................................................................. 89
7.3
INTERNATIONAL FRAMEWORK............................................................................................................................................ 93
7.4
STRATEGY ............................................................................................................................................................................... 95
7.5
CONCLUSIONS AND PERSPECTIVES ...................................................................................................................................... 96
REFERENCES .......................................................................................................................................................................................... 96
GLOSSARY OF ACRONYMS ............................................................................................................................................. 97
3
Introduction
Introduction
Nicoletta Calzolari, Claudia Soria
The context
Language Technologies (LT), together with their backbone, Language Resources (LR), provide
an essential support to the challenge of Multilingualism and ICT of the future. The main task of
language technologies is to bridge language barriers and to help creating a new environment
where information flows smoothly across frontiers and languages, no matter the country, and
the language, of origin.
To achieve this goal, all players involved need to act as a community able to join forces on a set
of shared priorities. However, until now the field of Language Resources and Technology has
long suffered from an excess of individuality and fragmentation, with a lack of coherence
concerning the priorities for the field, the direction to move, not to mention a common
timeframe.
This lack of coherent directions is partially also reflected by the difficulty with which
fundamental information about LR&Ts can be reached: basically, it is very difficult, if not
impossible, to get a clear picture of the current situation of the field in simple terms such as who
are the main actors, what are the available development and deployment methods, what are the
“best” language resources, what are the areas for which further development and investment
would be most necessary, etc. Substantial information is not easily reachable not only for the
producers but also for policy makers and funding agencies.
Under this respect, since some time large groups have been advocating the need of a LR&T
infrastructure, which is increasingly recognised as a necessary step for building on each other
achievements, integrating resources and technologies and avoiding dispersed or conflicting
efforts. A large range of LRs and LTs is there, but the infrastructure that puts LR&Ts together
and sustains them is still largely missing; interoperability of resources, tools, and frameworks
has recently come to be understood as perhaps the most pressing current need for language
processing research. Infrastructure building is thus indicated by many as the most urgent issue
and a way to make the field move forward, together with real sharing of resources, and an effort
to make resources available for all languages.
The context encountered by the FLaReNet project was thus represented by an active field
needing a coherence that can only be given by sharing common priorities and endeavours.
FLaReNet has contributed to the creation of this coherence by gathering a wide community of
experts and making them participate in the definition of an exhaustive set of recommendations.
What are Language Resources
The term “Language Resources” traditionally refers to usually large sets of language data and
descriptions in machine readable form, to be used in building, improving or evaluating natural
language, speech or multimodal algorithms or systems. Typical examples of LRs are written,
spoken, multimodal corpora, lexicons, grammars, terminologies, multimodal resources,
ontologies, translation memories, but the term is also extended to include basic software tools
for their acquisition, preparation, annotation, collection, management and use. The creation and
use of these resources span several related but relatively isolated disciplines, including NLP,
information retrieval, machine translation, speech, multimodality. FLaReNet endorses this
broader definition of LRs, having recognised the need for an “extension” of the term and in the
light of recent scientific, methodological-epistemological, technological, social and organisational
developments in the application fields of content processing/access/understanding/creation,
Human-Machine, Human-Human & Machine-Machine communication, and the corresponding
4
Introduction
areas from which the theoretical underpinnings of these application fields emerge (linguistics,
cognitive science, AI, robotics).
The FLaReNet mission, policy and strategy
FLaReNet – Fostering Language Resources Network – is an international Forum, composed by a
large community, aiming at developing the needed common vision and fostering a European
strategy for consolidating the sector, thus enhancing competitiveness at EU level and worldwide.
Its goals are:
• to facilitate interaction among LR&T stakeholders and ultimately re-build a community
around LR and LTs;
• to promote and sustain international cooperation;
• to coordinate a community-wide effort to analyse the sector of LR&Ts along all the
relevant dimensions: technical and scientific, but also organisational, economic, political,
cultural and legal;
•
•
to identify short, medium, and long-term strategic objectives and provide consensual
recommendations in the form of a plan of action targeted to a broad range of
stakeholders, from the industrial and scientific community to funding agencies and
policy makers;
to pave the way to the set up and functioning of an Open Resource Infrastructure
through a number of preparatory initiatives.
The FLaReNet community
FLaReNet brings together leading experts of research institutions, academies, companies,
consortia, associations, funding agencies, public and private bodies both at European and
international level. This way, we ensure that both LR producers and users are represented and
actively involved, in addition to technology developers. And in turn, broad participation of
different key-figures and experts ensures that recommendations are formulated through a
consensual bottom-up process in which the relevant scientific, technical, organisational,
strategic aspects and positions are taken into account.
The FLaReNet Community is composed of a network of 99 Institutional Members and 365
Individual Subscribers, representing 33 different countries. Community members belong to
academia, industry, resource distribution agencies, media, publishing houses, and national and
European institutional bodies. FLaReNet has also established a Network of National Contact
Points, for the purpose of obtaining up-to-date and reliable information about current initiatives
worldwide in the area of Language Resources (Data, Tools, Evaluation and Meta-Resources). The
Network consists at present of 105 experts from 34 countries or regions in the European Union
(26 Member States and 6 regions), 9 non-EU European countries and non-European countries
(36 countries). The role of the National Contact Points is to provide information on the
initiatives on Language Resources in their country/region, as well as to ensure a permanent link
between FLaReNet and their country/region, and to provide further reliable information about
the HLT research activities and entities, industries and administrations in their country/regions.
A Map showing all the countries for which there is a FLaReNet contact point is available on the
FLaReNet web site 1.
Over the duration of the project, FLaReNet has truly become a community of experts that – by
acting as an aggregator and facilitator of information sharing and discussion – is one of the first
scientific social networks in the field of Language Resources and Technology. This community
represents now a legacy that needs to be sustained to keep the momentum created by FLaReNet.
1
http:/www.flarenet.eu
5
The FLaReNet data
Introduction
Having recognised the lack of information about existing language resources as one of the major
factors hindering the development of the field, FLaReNet has undertaken a number of actions to
survey existing resources, inform about them, and enhance their visibility. Community
involvement is the underlying leitmotif that is common to all activities.
FLaReNet, together with ELRA, has launched the LRE Map of Language Resources and Tools. The
Map is an entirely new instrument for discovering, searching and documenting language
resources, here intended in a broad sense, as both data and tools. The purpose of the LRE Map is
to complement ongoing cataloguing efforts in order to enhance visibility for all resources, for all
languages and intended applications.
Other data come from the Repository of Standards, Best Practices and Documentation, the Wiki
Survey of the national and transnational initiatives in the area of Language Resources 2, the
Language Library, 3 a Database of the various technical and organisational methods, techniques
and models used by academic and industrial players in the area of Language Resources for
building and maintaining resources, and a Survey of methods for the automatic construction of
LRs. All these data are collected in the FLaReNet “DataBook”, a companion document to this
deliverable.
The FLaReNet recommendations: a collective enterprise
As a response to its main mission, FLaReNet has identified a number of consensual priorities and
strategic objectives for the field of Language Resources and Technology, which have been
provided to the community through a series of Blueprints of Actions and Infrastructures. The
whole community at large has been involved in this task, by means of a continuous direct
consultation of key players and stakeholders, in the attempt to reflect the widest possible view.
Moreover, FLaReNet has maintained regular cooperation activities with relevant projects and
initiatives, both non-European and European, in order to establish a kind of global coordination
of the LR field.
The FLaReNet recommendations target a broad spectrum of users:
• HLT stakeholders at large, including producers, users and developers of Language
Resources and Technologies, both academic and industrial (for instance, academic or
industrial researchers, service and media providers, providers of translation and
localization services, etc.)
• Funding agencies and policy-makers, at national and EC level.
Organisation of the Book
This volume collects and organises the major high-level recommendations collected around the
FLaReNet meetings, panels and consultations, as well as the results of the surveying and analysis
activities carried out under FLaReNet work packages. The FLaReNet Steering Committee took
care of summarising and organising the various recommendations, also as part of the work
carried out in the individual Work Packages of the project. As such, this Book is the result of a
permanent and cyclical consultation that FLaReNet has conducted inside the community it
represents – with more than 300 members – and outside it, through connections with
neighbouring projects, associations, initiatives, funding agencies and government institutions.
2
3
6
http://www.flarenet.eu/?q=WG7
For a description of the goals of the Language Library, see http://www.lrec-conf.org/lrec2012/
Introduction
The different chapters look at the various dimensions that are of relevance in the field, thus
providing different viewpoints on the topic of language resources and complementing each
other. They start from the current situation to identify areas for future developments.
This is why the reader will find some redundancy in the content of the chapters. This is done on
purpose as some issues, such as evaluation, standards, reuse etc. encompass various dimensions
and are therefore tackled in the different chapters from different point of views.
The Blueprint, in Chapter 1, provides a comprehensive perspective, highlighting the synergies
among the several positions, and summarising the fundamental recommended actions for the
development and progress of LRs and LTs, coming from the FLaReNet community and
addressing both the community itself and policy makers.
Chapter 2 looks at LR sustainability and maturity on the basis of a descriptive model of
sustainability from a LT perspective. It summarizes the description of a sustainability model,
elaborated within the project. Chapter 3 focuses on models for reusing and interlinking existing
resources and tools as a means to derive, produce and repurpose new LRs (datasets and
relevant processing tools). Chapter 4 focuses on the major conditions and recommendations
necessary to build the so-called “Interoperability Framework”, as a dynamic environment of
language (and other) standards and guidelines, where the former can interrelate coherently
with one another and the latter describe how standards work and “speak” to each other. Chapter
5 looks at issues of Language Resource quality in terms of validation and evaluation, and
describes and clarifies recommendations for these activities. Chapter 6 establishes some
recommendations for strategic actions to be put in force either by the community or by policy
makers towards an extensive automation of LR production, which will lead to beneficial
improvements for the commercial value of LTs and thus will enforce competitiveness of LTbased enterprises globally. Finally, Chapter 7 presents a strategy to fill in the gaps detected by a
survey on the national and cross-border R&D effort in language resources and technologies, and
existing data, tools, standards and evaluation methods.
Acknowledgements
This work could not have been possible without the invaluable contribution of all the people
who voluntarily shared their knowledge and understanding of the LRT landscape. We also
acknowledge Andrew Joscelyne for professional revision and editing of early drafts of this
document.
7
Ch 1. The Blueprint
Chapter 1. The FLaReNet Blueprint of Actions and Infrastructures
Nicoletta Calzolari, Valeria Quochi, Claudia Soria
with the contribution of Núria Bel, Gerhard Budin, Khalid Choukri, Joseph Mariani, Monica
Monachini, Jan Odijk, Stelios Piperidis
This chapter presents the final Blueprint of Actions and Infrastructures produced by the
FLaReNet project and community. The Blueprint is placed as the first chapter of the book as it
summarises and prioritises the contents of the other chapters. It is however conceived as an
autonomous document that provides the fundamental recommended actions for the
development and progress of LRT in Europe, coming from the FLaReNet community and
addressing both the community itself and policy makers.
The content of the Blueprint is a synthesis of the discussions and activities of the three years of
the FLaReNet project, and as such it integrates the previous two Blueprints 1 as well as the
outcomes of the stimulating discussions held both within and outside FLaReNet. Thanks to the
participation of outstanding experts in the field, the content of this document can be considered
as the expression of the Language Resources community setting its own agenda of future actions
to be undertaken concerning a broad spectrum of dimensions related to Language Resources 2.
Recognizing that the development of the sector of LTs is conditioned by various factors, all
interested stakeholders need to operate seriously together and forge partnerships to push LTs.
FLaReNet tackled this issue by bringing different stake-holders together, and having them
discuss about several key topics, such as the lack of infrastructures for the domain, and try to
foster joint plans, projects and roadmaps. Some early results are already visible: an EC funded
Network of Excellence started in 2010 – META-NET – with one of its main goals being the design
and set-up of an infrastructure for sharing language resources and technology at large – i.e.
META-SHARE. This was the main recommendation of FLaReNet in its first year.
Together, and under the umbrella of a shared view of today’s priorities, a future can be shaped in
which full deployment of Language Resources and Technologies is consolidated through
coordination of programs, actions and activities. While there has been considerable progress in
technology developments in the last decade, the significant challenge of overcoming current
fragmentation and imbalance inside the LTs community for all languages still remains an issue.
Thanks to initiatives such as the FLaReNet project, this situation is now starting to be tackled
and a new awareness is now spreading about the need and importance of joining forces and
build a compact community.
The FLaReNet recommendations cover a broad range of topics and activities, spanning over
production and use of language resources, licensing, maintenance and preservation issues,
infrastructures for LRs, resource identification and sharing, evaluation and validation,
interoperability and policy issues.
In principle, the addressees of this Blueprint belong to a large set of players and stakeholders in
Language Technologies (LTs), ranging from individuals to research and education institutions, to
policy-makers, funding agencies, SMEs and large companies, service and media providers. Its
main goal is thus to serve as an instrument to support stakeholders in planning for and
addressing the urgencies of the LRTs of the future. The recommendations contained in the
present document should therefore be taken into account by any player, whether on a European,
1
2
8
http://www.flarenet.eu/?q=Deliverables
According to the FLaReNet definition, Language Resources are all language data sets and basic tools, see the
Introduction.
Ch 1. The Blueprint
National, local, or private level, wishing to draft a program of activities for his/her own
communities.
How to use the Blueprint
In this document, the various actions recommended are organised around nine dimensions that
are relevant for the field of Language Resources: a) Infrastructure, b) Documentation, c)
Development, d) Interoperability, e) Coverage, Quality and Adequacy, f) Availability, Sharing and
Distribution, g) Sustainability, h) Recognition and i) International cooperation. Some of these
dimensions are of a more infrastructural nature, some are more related to research and
development, some yet more to political and strategic aspects, but they all must be seriously
considered when making up a strategy for the future of the field. All of them eventually have an
impact in the development and success of LRs, and represent the areas where actions need to be
taken to make the field of Language Resources and Technologies grow.
It is useful to see the various dimensions as a coherent system where each one presupposes the
other, so that action at one of the levels requires some other action to be taken at another one.
For instance, open availability of data presupposes interoperability (which in turn is boosted by
openness); to discover and develop new paradigms, and for data to be usefully exploited, the
availability of large quantities of data requires the ability to link the information carried by data.
Increased data quantity implies a change in their availability towards openness, etc.
Taken together, these directions are intended to contribute to the creation of a sustainable LRT
ecosystem.
1.1 Recommendations at a glance
The first section of this chapter presents an overview of the main recommendations in a
synoptic format, organised by dimensions. Each issue is then expanded and described in the
following sections. Each dimension is given a prototypical challenge, representing the “vision” or
condition to be attained according to the point of view of the FLaReNet community of experts.
The dimension “Infrastructure of Language Resources” is given a prominent and separate role in
that it encompasses most of other dimensions.
Recommendations are targeted to Language Resource Producers (LRP, broadly encompassing
academic and industrial developers) and/or Policy Makers (PM, i.e. National or supra-national
funding agencies, politicians, etc.).
An Infrastructure of Language
Resources
Dimension
Challenge
Recommended actions
Target
Build and sustain
the proper
Language
Resource
Infrastructure
• Build a sustainable facility for sharing
resource data and tools
• Establish international hub of resources and
technologies for speech and language
services, by creating a mechanism for
accumulating speech and language resources
together with industries and communities
• Develop and propose (free) tools and more
generally Web services (comparable to the
Language Grid), including evaluation
protocols and collaborative workbenches in
the LR infrastructure
LRP/PM
LRP
LRP
9
Ch 1. The Blueprint
Challenge
Recommended actions
Ensure that
Language
Resources are
accurately and
reliably
documented
• Devise and adopt a widely agreed standard
documentation template for each resource
type, based on identified best practice(s)
• Ensure that appropriate metadata are
consistently adopted for describing LRs
• Set up a global infrastructure of common and
uniform and/or interoperable metadata sets
• Develop and support community-wide
initiatives such as the LRE Map
• Establish links with other communities to
gain access to better information on the
existence of LRs in their domains, and
exchange Best Practices on handling and
sharing resources
• When producing a LR, allocate time and
manpower to documentation from the start;
provide documentation (or links to it) when
giving access to a LR
• Ensure strong public and community support
to definition and dissemination of resource
production best practices
• Go Green: enforce recycling, reusing and
repurposing
• Encourage the full automation of LR data
production
• Invest in Web 2.0/3.0 methods for
collaborative creation and extension of highquality resources, also as a means to achieve
better coverage
• Start an open community initiative for a large
Language Knowledge Repository
• Estimate the cost of producing LRs needed to
develop an LT for one language
Documentation
Dimension
Development
Define a reference
model for future
LR development
10
Target
LRP
LRP
LRP
LRP
LRP
LRP
PM
LRP
LRP
LRP/PM
LRP
LRP
Ch 1. The Blueprint
Coverage, Quality, Adequacy
Interoperability
Design and set up
an interoperability
framework for
LRs and LT
Address
appropriate
coverage in terms
of quantity, quality
and adequacy to
technological
purposes
• Ensure formal and semantic interoperability
of Language Resources
• Identify new mature areas for
standardisation and promote joint efforts
between R&D and industry
• Make standards operational and put them in
use
• Invest in standardisation activities
• Encourage/enforce use of best practices or
standards in LR production projects
• Set up an “interoperability challenge” as a
collective exercise to evaluate (and possibly
measure) interoperability
• Collaboratively build and maintain a
repository of standards and best practices,
linked to standards-compliant open/freely
available data
• Create a permanent Standards Observatory
or Standards Watch
• Set up training initiatives to promote and
disseminate standards to students and young
researchers
• Increase quantity of resources available to
address language and application needs
• Implement BLaRKs for all languages,
especially less-resourced languages
• Address formal and content quality of
resources by promoting evaluation and
validation
LRP
LRP
LRP
PM
PM
LRP
LRP/PM
PM
LRP/PM
LRP
LRP
PM/LRP
• Establish a European evaluation and
validation body
PM
• Assess availability of resources with respect
to their adequacy to applications and
technology requirements
LRP/PM
• Define and establish a Quality Certificate or
quality score for LRs, to be endorsed by the
community
LRP
11
Availability: Sharing and Distribution
Sustainability
Recognition
International Cooperation
12
Make resources
easily and readily
available, within
an adequate IPR
and legal
framework
Ensure
sustainability of
language
resources
Promote the LR
ecosystem
Promote synergies
among initiatives
at international
level
Ch 1. The Blueprint
• Opt for openness of LRs, especially publicly
funded ones
• Ensure that publicly funded resources are
publicly available either free of charge or at a
small distribution cost
• Create LRs in collaborative projects where
resources are exchanged among project
participants after production
• Make LRs available for most types of use,
making use of standardised licenses where
reuse is clearly stipulated
• Educate key players with basic legal know
how
• Elaborate specific, simple and harmonised
licensing solutions for data resources
• Clear IPR at the early stages of production;
try to ensure that re-use is permitted
• Make sharing/distribution plans mandatory
in projects concerning LR production
• Ensure that all LRs to be produced undergo a
sustainability analysis as part of the
specification phase
• Assess LR sustainability on the basis of
impact factors
• Foster use of a sustainability model
• Develop a standard protocol for citing
language resources
• Give greater recognition to successful LRs and
their producers (Prizes, Seals of Recognition,
etc. )
• Support training in production and use of LRs
• Share the effort for the production of LRs
between international bodies, such as the EC
for the EU, and individual countries/regions,
such as Member States in Europe
• Encourage agencies to open their calls to
foreign participants
• Establish MoU among existing initiatives,
starting with the EC efforts
• Ensure the sustainability of the FLaReNet
International Contact Points, also through the
Survey Wiki
• Produce a White paper on LT, including LR, in
preparation for FP8. Distribute it within
Europe and abroad
LRP/PM
PM/LRP
LRP
LRP
LRP/PM
LRP
LRP
PM
PM/LRP
LRP
LRP/PM
LRP/PM
LRP/PM
LRP
PM
PM
PM
PM
LRP
Ch 1. The Blueprint
1.2
Infrastructure of Language Resources
“Our vision is a scientific e-Infrastructure that supports seamless access, use, re-use and trust of
data. In a sense, the physical and technical infrastructure becomes invisible and the data
themselves become the infrastructure – a valuable asset, on which science, technology, the economy
and society can advance (High Level Expert Group on Scientific Data, 2010)”
Build a sustainable facility for sharing resource data and tools
Make LRs available, visible and easily accessible through an appropriate
infrastructure; participate in building the infrastructure by providing feedback on
relevant actions in this direction
Ensure a stable sustainable infrastructure for LR sharing and exchange; support
continuous development and promotional activities
Establish an infrastructure that can help LR producers with legal issues
Support the emergence of a tool-sharing infrastructure to lower the cost of R&D
for new applications in new language resource domains
Establish international hub of resources and technologies for speech and
language services, by creating a mechanism for accumulating speech and
language resources together with industries and communities
Develop and propose (free) tools and more generally Web services (comparable
to the Language Grid), including evaluation protocols and collaborative
workbenches in the LR infrastructure
LRP
PM
•
•
•
•
•
•
•
•
•
•
The need for an infrastructure for Language Resources was the first recommendation since the
beginning of FLaReNet and derives historically from the recognition of the infrastructural role of
LRs as essential “building blocks” for language technologies.
Such an infrastructure should ease recovery and use of LRs through appropriate facilities that
allow their availability, visibility, and easy accessibility.
A small number of infrastructure initiatives (cf. CLARIN, META-SHARE) have emerged in order
to solve these problems. However, without proper coordination between them, there is a risk of
further fragmentation. An additional problem is that most of the current “infrastructure plans”
are projects with limited time duration. It is therefore time to act more decisively to build
synergy between all stakeholders in this field.
The basic principles of an infrastructure for language resources and technologies require a
community approach that brings together and builds on current experiences and endeavours. It
is necessary to define and agree on the basic criteria and dimensions for an appropriate
governance, and define the basic data and software resources that should populate this
infrastructure. Multilingual coverage, the capacity to attract providers of useful and usable
resources, improvements in sharing mechanisms, and collaborative working practices between
R&D and commercial users are key aspects. There must also be a business-friendly framework
to stimulate the commercial use of these resources, based on a sound licensing facility, ease of
access, ease of conversion into uniform formats.
As a matter fact, researchers and developers still have to spend quite some time consulting
multiple catalogues or searching the web to find relevant LRs, and often they fail. Available
resources are very often difficult to access for various reasons. Some are available from
distribution centres (notably ELRA and LDC), others from portals of projects or associations, or
directly from the web pages of the laboratories or researchers who developed the resource, or
even from the owner itself. In many cases, unless the potential user already knows something
13
Ch 1. The Blueprint
about the resource (s)he might want to use (e.g. name, owner, or project), (s)he would find it
difficult to discover new or as yet unknown resources. The identification and discovery of
“new/unknown” resources is therefore a key priority right now and should be accompanied by
the spread of a new culture of sharing and/or collaborative resource creation (see below, 1.4
and 1.7).
These basic principles for LR infrastructures are related to or incorporated in most of the issues
appearing in the following sections, and will therefore not be expanded here.
1.3
Resource Documentation
“Ensure that Language Resources are accurately and reliably documented”
LRP
Devise and adopt a widely agreed standard documentation template for each
resource type, based on identified best practice(s)
•
Ensure that appropriate metadata are consistently adopted for describing LRs
•
Ensure strong public and community support to production and dissemination of
best practices
Provide appropriate MD description for all LR distributed, preferably in one of
the widely used MD schemes
Support MD creation and promotion activities; set guidelines and rules for MD
description of available LRs and support relevant efforts
Create machine understandable MD with formal syntax and clear semantics
Use formalized MD elements as much as possible
Put all aspects of the documentation of the LR that can be formalized in
formalized metadata elements
Ensure that all data categories used in the metadata and in the data are
registered in a data category registry to ensure semantic interoperability
Automate the process of MD creation
Set up a global infrastructure of common and uniform and/or interoperable
metadata sets
Agree on a limited/basic set of MD
Develop structured MD
Develop and support community-wide initiatives such as the LRE Map
Establish links with other communities to gain access to better information on
the existence of LRs in their domains, and exchange Best Practices on handling
and sharing resources
When producing a LR, allocate time and manpower to documentation from the
start; collect and provide documentation (or links to it) when giving access to a
LR
In a LR production project, part of the funding should be allocated to
documentation and dissemination activities; support activities for collecting and
storing in appropriate infrastructures documentation for LRs
•
•
PM
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Accurate and reliable documentation of Language Resources is an undisputable need. Instead, as
of today, LRs are still often poorly documented or not documented at all, and, even when
available, documentation is often not easy to find.
14
Ch 1. The Blueprint
Documentation allows language resources to be used by people different from those who
designed and developed them. The variable nature of documentation can hamper the
dissemination and replication of LRs and makes it hard for users to read and compare how-to
files. Common best practices for documentation and guidelines writing need to be established
and enforced. This entails developing standard specifications for LR documentation.
Documentation should be as exhaustive as possible, and include information about data format
and data content, the production context, and existing possible applications.
What users need is information that helps them:
a) Find a resource and assess its usefulness for a
given application
b) Understand the production process, the use of
“Researchers
and
best practices, and intended exploitation
practitioners will be able to
c) Assess the quality of a resource
find, access and process the
d) Replicate processes and results
data they need. They will be
e) Handle idiosyncrasies or documented errors.
Machines need (machine-understandable) information
confident in their ability to
to:
use and understand data and
they can evaluate the degree
a) Discover and compare resources
to which data can be trusted”
b) Validate formats and annotations
(D. Giaretta, 2011)
c) Process annotations appropriately
d) Retrieve relevant parts of a resource for a given
use
e) Enable other as yet unexplored new uses.
Guidelines are also useful for replicating and extending resources. They may re-channel
language resource efforts in a more coherent way (see the experience in SPEECHDAT 3) and help
develop similar resources for other languages well past the end of a given project. There should
therefore be a set of exhaustive and reliable guidelines for every resource type, building on the
experience of successful projects.
An effort must be made to collect all existing LR documentation and make it easily available. To
this end, the design and construction of a (virtual) repository of specifications, guidelines, and
documentation of LRs, starting with reference resource models or widely known and used
resources (e.g. WordNet, Penn TreeBank, …) is a priority task 4.
Uniform documentation is vital. A common documentation template should be defined,
promoted, and enforced for all contracts for publicly funded projects.
Documentation should include:
a) A high-level description giving the non-expert but interested reader a good idea of what
is in the resource, including general information such as owner/copyright holder, format
and encoding of the data and the files, languages(s), domains, intended applications,
applications in which the data has been used, and details about basic quality assessment
(in particular for availability/reliability of the encoded information).
b) Information on the theoretical framework, background, and/or the “philosophy” of the
resource.
c) Specifications of the methodology used to create the resource so that others could
replicate the process.
d) Annotation specifications (with data categories and their semantics) and guidelines, i.e.
guidelines used by annotators.
3
4
http://www.speechdat.org
An activity along these lines has been started within FLaReNet, see
http://www.flarenet.eu/?q=FLaReNet_Repository_of_Standards_and_Guidelines
15
Ch 1. The Blueprint
e) Information on the use of standards (at all levels: production, annotation, validation,
etc.).
f) Specification of the methodology or guidelines used to assess the quality of the resource
(if validation is conducted) and the report on such validation.
g) Estimates of the effort required to create the resource (in any reproducible unit, e.g.
person/month).
Documentation is also the gateway to LR discovery. Ensuring that Language Resources are
discoverable is the first step towards promoting the data economy. In order to make LRs
discoverable, language resource providers (LRPs) should always document their resources,
using standard metadata and unique resource identifiers. A useful best practice in this line is to
document one’s resource in an open catalogue, such as the ELRA Universal Catalogue 5, the
META-SHARE 6 or CLARIN 7 catalogues. By providing standard APIs to query the catalogues, it
would then be possible to harvest individual catalogues and compile unified catalogues by
metadata aggregation.
One of the main reasons why it is now difficult to find resources that match specific needs and
languages is the lack of compatibility for metadata. Different sub-communities, data distribution
centres, archiving institutions and projects, and other providers tend to use their own, noninteroperable metadata sets to describe their data. Resources are also described at different
levels of granularity depending on who does it. Another drawback is that it is often impossible to
combine data from multiple sources to create new data sets for specific uses. We need a solution
to the lack of a “one-size-fits-all” resource.
In this context, standardised metadata play a crucial role, and their use should be made
obligatory in all resource production projects. The key priority is therefore to work towards the
full interoperability of metadata sets. As there are many differing metadata sets and search
engines, harmonisation is a central problem for the community. Useful initiatives in this
direction are community-based documentation initiatives, such as the LRE Map 8, by which
massive documentation of existing resources is achieved in a limited time frame and with
limited effort, with the additional advantage that all resources are documented in a uniform and
standard-compliant way. Therefore, definition and adoption of standardised metadata must be
the first priority and first step.
It is important to shift the metadata mind-set towards the creation of machine-understandable
metadata, i.e. pieces of information about (digital) resources that can be processed automatically
(e.g. they must have a formal syntax and a declared semantics). This will make metadata
browsable/accessible from various tools for various purposes. The development of techniques
for automating the process of metadata creation would also help spread the adoption of
machine-understandable metadata.
5
6
7
8
http:// universal.elra.info/
http:// www.meta-net.eu/meta-share/meta-share
http:// www.clarin.eu/vlo/
http://www.resourcebook.eu/
16
Ch 1. The Blueprint
1.4
Resource Development
“Define a reference model for future LR development”
LRP
PM
Ensure strong public and community support to definition and dissemination of
resource production best practices
Ensure that all "best/good practices" are disseminated to the R&D community
through very lightweight methods, instead of the heavy machinery often used
Go Green: enforce recycling, reusing and repurposing
Implement tools that can help crawl “comparable” resources, select a small set
that can be aligned and then provide clear legal conditions to ensure they can be
shared with the MT community and beyond (assuming these can be annotated for
bilingual lexical extraction, named entities, etc.)
•
•
•
Promote more efficient uses of available data through a better grasp of
repurposing and merging techniques
•
Adapt LT from the language in which it exists to a similar one in the same
language family, then improve the quality of the LRs for the specific data in that
similar language through bootstrapping
•
Promote activities that can contribute to the (automatic) production of large scale
and high quality resources
•
Investments in research on the automatic production of language resources
should be increased to broaden the range of language resources addressed
•
Invest in new methods such as active learning and semi-supervised learning that
•
Analyse the relationship within language families and use joint LR methods to
develop LT within those families. Consider the use of pivot languages within
those families, based on the existence of parallel corpora
Encourage the full automation of LR data production
Develop a certain volume of MT-based corpora, post-edit them
(semi)automatically to specific quality levels, and then use them iteratively to
train new MT systems more effectively
Support academic and industry involvement in research on automatic methods
for production and validation of LRs, to allow a more accurate assessment of the
automatic methods for building LRs for real-world applications
Promote research into new methods and techniques for collecting information
from unstructured (non-annotated) data
•
•
Check the availability of existing LRs; try to reuse what is available; check what is
available not only for one's own language, but also for languages of the same
family; try to see what can be reused in terms of data and tools but also
experience; create and exchange tools that help in
reuse/repurposing/interlinking
support reuse projects and research activities in the domain; support activities
that document reuse, formulate guidelines, etc.; fund promotional and research
activities around reuse (e.g. dedicated workshops, publications on successful
reuse cases etc.)
•
•
•
•
•
•
•
•
•
•
•
17
deliver promising results for reducing the need for human intervention
Ch 1. The Blueprint
Use automatic content extraction from the Web, but bear legal issues in mind
Conduct a study and propose a Code of Ethics (regarding privacy and IPR) for
automatic information extraction on the internet
Develop tools for anonymizing data
Invest in Web 2.0/3.0 methods for collaborative creation and extension of highquality resources, also as a means to achieve better coverage
Carry out LR production and annotation as collaborative projects; open up
existing LRs for collaborative annotation and the reuse of the annotated results;
participate in communities that carry out similar tasks; evaluate and document
their results; develop new tools and/or adapt existing tools and platforms to the
needs of collaborative work
Support collaborative efforts for large LR production and annotation projects
through appropriate funding; support the infrastructure required for
collaborative work
•
•
•
•
•
•
Promote the collaborative development of
semantically/pragmatically/dialogically annotated corpora
•
Foster the debate and experiments on new outsourcing trends over the web
•
Propose a Code of Ethics for crowdsourcing, in order to guarantee a sufficient
salary and compliance with labour rights
Start an open community initiative for a large Language Knowledge Repository
Estimate the cost of producing LRs needed to develop an LT for one language
•
•
•
•
•
•
Development of language resources refers to the entire production cycle of a resource.
The proper management of the “life cycle” of language resource creation has attracted less
attention has and been largely overlooked in our community. A reference model for creating
Language Resources instead will help address the current shortage of resources in terms of
breadth (languages and applications) and depth (data quality and volume). Such reference
model should also include an accurate estimate of the production costs.
The creation of new resources from scratch should be discouraged wherever resources can be
found for a given language and/or application. We should encourage re-use and re-purposing via
a “recycling” culture to ensure the reuse of development methods, existing tools, and
translation/transliteration tools, etc. (see Ch. 3).
The experience gained for one language can be used to
process others. It is encouraging to see high-level
“Innovation needs data, but
applications for Less-Resourced Languages (instead of just
also the collection of data
the usual “taggers”) such as ASR for Amharic, as these can
needs innovation”
pave the way for designing baseline systems for these
(Vasiljevs, 2011).
languages.
Similarly, most language technologists use existing
language resources as input and create content as byproducts that could form useful language resources for others. Yet so far very few of these
resources are made commonly available at the end of the production cycle. With production
costs constantly increasing, there is a need to invest in innovative production methods that
massively involve automatic procedures, so as to reduce human intervention to a minimum (see
also Ch. 6).
18
Ch 1. The Blueprint
The coverage problem is so enormous that we must promote strategies that approach or ensure
full automation for (high-quality) LR data production. We must improve existing tools and
introduce new automation techniques, especially for higher-level semantic, content-related and
multilingual tasks. We must also foster the evaluation of real-life applications so that research
can gradually approach industry needs in terms of information volume and granularity.
Given the high cost of language resource production, and that in many cases it is impossible to
avoid the manual construction of resources (e.g. if accurate models are requested or if there is to
be reliable evaluation) it is worth considering the power of social/collaborative media to build
resources, especially for those languages where there are no language resources built by experts
yet.
There are several experiments in crowd-sourcing data collection and NLP tasks (Dolan, 2011),
and most of them look promising. Crowd-sourcing (such as Amazon’s Mechanical Turk) makes it
possible to mobilize large armies of human talent around the world with just the right language
skills so that it is feasible to collect what we need when we need it, even during a crisis such as
the recent earthquake in Haiti or the flood in Pakistan (Church, 2011). For instance, it has been
estimated that Mechanical Turk translation is 10 to 60 times less expensive than professional
translation (Callison-Bruch and Dredze, 2010).
However, the use of crowd-sourcing raises ethical, sociological and practical issues for the
community. It is not yet clearly understood for example whether all types of LRs can be obtained
collaboratively by using naïve annotators; more research is therefore needed on both the
technical (e.g. accurately comparing the quality and content of resources built collaboratively
and those built by experts) and ethical aspects of crowd-sourcing 9 (see for instance Zaidan and
Callison-Burch 2011 about mechanisms for increasing quality of crowd-sourced data).
A particularly sensitive case is that of less-resourced languages, where language technology
should be developed rapidly to help minority-language speakers access education and the
Information Society (see also Ch. 7). Basic language resources for all the worlds’ languages could
be created building a Web 2.0 site (using the same community computing power that generates
millions of blogs) starting with the 446 languages currently present on the web. Collaborative
and Web 2.0 methods for data collection and annotation seem particularly very well-suited for
collecting the data needed for the development of LT applications.
There are insufficient current resources and sources to solve the problem of creating free, largescale resources for the world languages, even for those with a reasonable web presence. The
collaborative accumulation and creation of data appears to be the best and most practicable way
to achieve better and faster language coverage and in purely economic terms could well deliver a
higher return on investment than expected.
9
See for instance (Zaidan and Callison-Burch 2011) about mechanisms for increasing quality of crowd-sourced data.
19
1.5
Resource Interoperability
Ch 1. The Blueprint
“Design and set up an interoperability framework for Language Resources and Technology”
LRP
Ensure formal and semantic interoperability of Language Resources
Define the standard as the lowest common denominator, at the highest level of
granularity, a basic principle that was already factored into EAGLES
Standards must generalize over multilingual data and should be tested on
multilingual data
Identify new mature areas for standardisation and promote joint efforts between
R&D and industry
Engage in groups that produce (new) standards and provide feedback and
comments
Make standards operational and put them in use
Encourage the building of tools that enable the use of standards, and step up the
availability of sharable/exchangeable data
Fund the development and/or maintenance of tools that
support/enforce/validate standards
Invest in standardisation activities
Support infrastructural activities for collecting and disseminating information on
existing standards and best practices
Fund activities for setting up new standards where they do not exist
Closely monitor the Linked Open Data initiative tightly connected to semantic
interoperability, so that we understand the potentialities of this initiative for our
field
Encourage/enforce use of best practices or standards in LR production projects
Look for standards and best practices that best fit the LRs to be produced, already
at the early stages of design/specifications; adhere to relevant standards and best
practices; produce LRs that are easily amenable to reuse (e.g. adopt formats that
allow easy reuse)
Create ‘official’ validators (such as http://validator.oaipmh.com/ or the OLAC
validator) to check compliance of LRs with basic linguistic standards
Set up an “interoperability challenge” as a collective exercise to evaluate (and
possibly measure) interoperability
Create a permanent Standards Observatory or Standards Watch
Set up training initiatives to promote and disseminate standards to students and
young researchers
PM
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Interoperability of resources is the extent to which they are formally and content compatible, so
as to allow, for instance, the merging of data coming from different sources while preserving
their semantics.
We can distinguish between syntactic and semantic interoperability. Syntactic interoperability is
the ability of different systems to process (read) exchanged data either directly or via trivial
conversion. Semantic interoperability is the ability of systems to interpret exchanged linguistic
information in meaningful and consistent ways via reference to a common set of reference
categories (Ide and Pustejovsky 2011)
20
Ch 1. The Blueprint
Today the lack of interoperability and compliance with standards costs a fortune. It is estimated
that buyers and providers of translation lose 10% to 40% of their budgets or revenues because
language resources are not stored in compatible standard formats (van der Meer 2011).
Interoperability of resources and data is also an essential prerequisite for successful exploitation
of the enormous amount of data that the advent of the Internet has been making available since
less than two decades. Data access and links within and across this data is as important as the
actual quantity, and data interoperability is essential to it.
While, on the one hand, it is increasingly recognised that standards are key to resource sharing,
re-usability, maintainability and long-term preservation, Language Resource Producers are still
largely lacking a clear understanding about why standards help represent data, and why there
are the advantages to adopting standards. As a result, many types of resources and many levels
of information and annotations are not standardised.
Most existing resources use unique representation formats and conventions, so other people
have to first understand the format and then build ad hoc conversions in order to use the
resource data for their own activities. This makes it especially difficult to draw on different
sources to build on-demand resources needed by emerging web technologies. The lack of
standardisation also makes it difficult to evaluate the quality and value of resources for a given
application. A basic level of standardisation is particularly vital for so-called “Less-Resourced”
Languages.
One solution would be to work towards the establishment of a broad-based framework for
interoperability of language resources and language technologies, involving industry in the mix.
There should be greater awareness of the importance of standards for resource
producers/managers who want to join the open-access club and boost the utilization of their
resources, so as to increase visibility, and attract more users and funding. Industry involvement
in standardisation initiatives will grant that there is broad-based adoption of standards.
An initiative towards these goals has been established by FLaReNet with a document (Calzolari
et al. 2011) proposing an overview of the current scene towards an Interoperability Framework
and acts as a reference point for the current standards encouraged by the community for
adoption 10.
The community and funding agencies need to join forces to drive forward the use of existing and
emerging standards, at least in the areas where there is some degree of consensus (e.g. external
descriptive metadata, meta-models, part-of-speech (POS) and morpho-syntactic information,
etc.). The only way to ensure useful feedback to improve and advance is to use these standards
on a regular basis. It will be thus even more important to enforce and promote the use of
standards at all stages, from basic standardisation for less-resourced languages (such as
orthography normalization, transcription of oral data, etc.) to more complex areas (such as
syntax, semantics, etc.).
However, enforcing standards cannot be a purely top-down process. It must be backed by
information about contributions from different user communities. As most users are not very
concerned about whether or not they are using standards, there should be easy-to-use tools that
help them apply standards while hiding most of the technicalities. The goal would be to have
standards operating in the background as “intrinsic” properties of the language technology or
the more generic tools that people/end-users use. The design of interoperability tasks will also
help to determine which emerging standards are most interoperable. Interoperability tests can
also replace aspects of validation (see Ch. 5).
At the same time, there should be a regular examination of new fields to check whether they are
“mature” enough to start a standardisation initiative (for instance about semantic roles and
spatial language). To this end a joint effort between academia and industry will again be
10
This initiative is in close synchronization with other relevant initiatives such as CLARIN, ELRA, ISO and TEI and
META-Share.
21
Ch 1. The Blueprint
advantageous and is thus to be promoted also in order to identify new areas that are mature for
standardisation activities.
1.6
Resource Coverage, Quality, Adequacy
“Address appropriate coverage in terms of quantity, quality and adequacy to technological
purposes”
With the current data-driven paradigm in force,
innovation in LT crucially depends on language resources
nowadays. Accent is being increasingly put on high quality
and huge size of resources, and as production (still) takes
a lot of effort and is very costly, development of the
resources for future technologies and applications must
start now in order to positively impact the development of
multilingual technologies such as Machine Translation,
cross-lingual and Web 3.0 applications.
Despite the vast amount of academic and industrial
investment, there are not enough available resources to
satisfy the needs of all languages, quantitatively and
qualitatively. Language resources should be produced and
made available for every language, every register, every
domain to guarantee full coverage, high quality and
adequacy for the various LT applications.
We need the right amount, the right type and the right
quality of resources.
1.6.1
“The development of killer
applications
crucially
depends on the availability of
large quantities of data.
knowledge
Cross-lingual
extraction, for instance, is a
challenging high impact task
for the near and mid future.
Today, the tasks seems to be
achievable because critical
mass of technology is
collected” (Mladenic and
Grobelnik, 2011)
Resource Quantity
LRP
PM
Increase quantity of resources available to address language and application
needs
•
•
Implement BLARKs for all languages, especially less-resourced ones
•
•
Enforce shared/distributed construction of resources as a means to achieve
better coverage
Analyse the dialectal variants of languages and include those variants in the
production of LRs
Provide sustainable high quality resources for all European languages
Ensure that resources are developed for all EU languages so that
evaluations can be made of language-dependent technologies listed by
the major agencies (training and testing data)
•
•
•
•
•
•
One thing that must be borne in mind is that dependence on data creates new disparities for
under-resourced languages and domains. It is estimated that 95% of web pages are in the top 20
languages (Pimienta et al., 2009). Naturally, smaller language communities produce much less
data than speakers of the languages dominating the globe. The same problems occur for
language data in narrow domains with their own specific terminological and stylistic
22
Ch 1. The Blueprint
requirements. Thus, provision of high quality resources for all European language, including
minority ones is a priority now, in order to avoid disparity in the future.
To ensure Universal Linguistic Rights and massive deployment of LT applications, language
services will need to be provided for everyone in their own mother tongue. This priority is also
evidenced by the number of localisation projects for most existing applications, be they
proprietary or open-source. As the quality of volunteer-based only localisation is not high,
funding must be found to cover all languages (including the world’s less-well represented
languages) in future multilingual applications by developing language resources for all
languages.
Specifically for the advancement of LTs, Basic Language Resource Kits (or BLARKs 11) should be
supported and developed for all languages and, at least, main applications (MT, IR, QA to
mention some). Also, as many of the undocumented languages of our cultural legacy may
become extinct in the digital age, minority and fringe languages should be comprehensively
represented through spoken and written corpora, and manuscripts should be digitized.
In this direction, first the BLaRK concept needs to be worked out in detail, so that it can be
embodied as a standard, and possibly planned revision sessions should be set, as it is
intrinsically a dynamic notion that changes in time with the change in technology development
in the different countries. Second, regular BLaRK surveys must be conducted to produce a clear
picture of technology trends, and establish (and regularly update) a roadmap covering all
aspects of LT. Third, resource production should be funded on the basis of BLaRK-like criteria,
i.e. giving priority to the development of “missing” resource types for each language.
Allocating funding to cover all languages (in particular less well represented ones) and all basic
needs of language technology remains thus a high priority for ensuring multilingual applications
in the future.
1.6.2
Resource Quality
LRP
Address formal and content quality of resources by promoting evaluation and
validation
Establish a European evaluation and validation body
Provide high-quality resources for all European languages
Promote automatic techniques to guarantee quality through error
detection and confidence assessment
Promote technologies that deliver highly accurate, high-confidence
results to reduce the amount of subsequent revision (human or
automated)
Establish common and standard Language Technology evaluation procedures
Organize (a significant part of) research around evaluation tasks,
challenges and campaigns
Promote evaluation as a driving force for research in HLT, in particular
for the development of techniques for fully automatic evaluation and of
fully formalized evaluation metrics
Devise new methods for LR quality check
Develop new and/or ensure the maximal use of existing tools for
11
http://www.blark.org/
PM
•
•
•
•
•
•
•
•
•
•
•
•
•
23
automatic or semi-automatic formal and content validation of LRs
Ch 1. The Blueprint
Create a think tank with recognized experts from a broad spectrum of the
community (academia/industry; technologies; and modalities
(written/speech/multimodal), etc.) to assess requirements for LR quality
Use crowd sourcing techniques to carry out validation tasks, but be
aware of the ethical and legal issues associated with these techniques
Create an infrastructure for coordinated LRTs evaluation
Secure an evaluation framework to assess the progress made by
technologies with respect to state of the art
Set up an evaluation management and coordination structure to ensure a
viable, consistent and coherent programme of activities that can
successfully scale up and embrace new communities and technological
paradigms
Set up a sustainable technical infrastructure providing data, tools and
services to carry out systematic evaluation. This could be a distributed
infrastructure involving existing organizations
•
•
•
Carry out evaluation in real-world scenarios
Define and establish a Quality seal of approval, on the model of the “Data Seal of
Approval” to be endorsed by the community
•
•
•
•
•
Provide funds for financing the evaluation management and coordination
structure and its associated technical infrastructure. This may require
new funding strategies and instruments both at the European and at
national and regional levels
Promote evaluation and validation activities of LRs and the dissemination of their
outcomes
•
•
•
•
•
High quality resources should be regarded as a key driver for effective technology in broad areas
(e-content, media, health, automotive, telecoms, etc.).
Evaluation in Europe is currently carried out by individual institutions (such as ELDA and
CELCT) and by short-term projects (e.g. the TC-STAR and CHIL campaigns), but there is no
sustained European-wide coordination, as there is in the US (NIST) or Japan (NII). In specific
areas, the community may organise itself to carry out regular evaluations (e.g. CLEF 2000-2010,
and Semeval) but with limited funding and much community good will.
In the US, however, NIST plays a very important role in coordinating technology evaluation as it
enables LRs to be controlled and streamlines the research and development of applications with
genuine commercial promise.
Evaluation should encompass technologies, resources, guidelines and documentation. But like
the technologies it addresses, evaluation is constantly evolving, and new, more specific
measures using innovative methodologies are needed to evaluate the reliability of semantic
annotations, for example.
Current evaluation campaigns sometimes create rather artificial settings so they stay
'academically clean', making the tasks they measure somewhat unrealistic. One of the most
critical challenges, therefore, is to introduce new types of campaigns, possibly based on taskbased evaluation. For practical purposes, it would be helpful to have guidelines and do's/don’ts
for this.
In order to foster evaluation activities, it would be important that they were highlighted as a
major research topic (which includes research on metrics, methodologies, etc.) especially as a
PhD subject.
24
Ch 1. The Blueprint
Thorough dissemination and information of activities and achievements should be done through
LRT evaluation portals (e.g. the ELRA HLT evaluation portal 12).
Evaluation packages should be distributed among the community or LT players.
New methods for validating and evaluating LRs (and LT in general) should be sought, proposed
and widely agreed.
Projects/players should be pushed to specify a validation procedure before LR production starts.
The community and policy makers should ensure that a quick quality check can be carried out, at
least for existing resources.
1.6.3
Resource Adequacy
Assess availability of resources with respect their adequacy to applications and
technology requirements
Assess maturity of technologies for which resources should be developed
Monitor ongoing research developments through publications and patent
filings, and keep track of prize-winning papers and theses
Draw a list of top 20 technologies and ensure that crucial resources are
produced in at least 10 of these, in a publicly and fully funded framework
Support EU-critical HTL domains one they have been identified
LRP
PM
•
•
•
•
Define a small set of crucial topics on which a critical mass of
research/researchers should focus through a clear adherence to its
principle
Introduce current industrial needs into the research agenda. Publish data
on whether resources are actually used, which aspects of LRs should be
made public, and help industry drill down into research on automatic
methods
Produce appropriate spoken language resources to study and develop a more
natural approach to voice input
•
•
•
Conduct regular surveys of technology transfer
Secure an evaluation framework to assess the progress made by
such technologies with respect to state of the art
Ensure regular evaluation campaigns to assess state of the art
•
•
•
•
•
•
•
•
Not only do we need more data, but the typology of data needed has also increased. “Nowadays
data can be emails, Facebook walls, and exchanges on Twitter. Today, data is gathered not only
from the Internet but also from supermarket receipts, mobile phones, cars, planes and soon even
refrigerators, ovens and any type of electronic device we use will provide data. Much of the data
that previously simply disappeared after having been used for a specific purpose, is now stored,
distributed and even resold for analysis, interpretation or other purposes of which the best if not
most frequent case is innovation” (Segond, 2011). However data access and links within and
across this data is as important as the actual quantity.
12
http://www.HLT-evaluation.org/
25
1.7
Resource Availability: Sharing and Distribution
Ch 1. The Blueprint
“Make resources easily and readily available, within an adequate IPR and legal framework”
By availability here it is intended the way in which a given resource is actually made available
for use by third parties. This implies decisions about licensing and business models.
Opt for openness of LRs, especially publicly funded ones
Ensure that publicly funded resources are publicly available either free of charge
or at a small distribution cost
Create LRs in collaborative projects where resources are exchanged among
project participants after production
Make LRs available for most types of use, making use of standardised licenses
where reuse is clearly stipulated
Educate key players with basic legal know how
Elaborate specific, simple and harmonised licensing solutions for data resources
Harmonise legislation regarding LR use for all types of LRs, and make provisions
for allowing the free use thereof at least for research/non-profit purposes
Develop clear agreement templates for sharing components and data, especially
to ensure optimal cooperation from commercial parties while safeguarding their
interests
Clear IPR at the early stages of production; try to ensure that re-use is permitted
Make sharing/distribution plans mandatory in projects concerning LR production
LRP
PM
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Data Openness
“Producers of data will benefit from opening it to broad access and will prefer to deposit their data
with confidence in reliable repositories (Giaretta, 2011)”
There is strong impulse towards open data nowadays, in the sense of data that are easily
obtainable and can be used with few, if any, restrictions. According to the Open Knowledge
Foundation, data is open if “you are free to use, reuse, and distribute it – subject only, at most, to
the requirement to attribute and share-alike”. The language resource community has started to
embrace this view and is inclined to think of open data as digital resources distributed under
open source-type licenses that can in turn be used, modified (and redistributed). While the
majority of LR experts advocate for data openly available and reusable, it is a fact that 55% of
the resources documented by the LRE Map are freely available. Reluctance in fully embracing an
open data model is still common. As pointed out by Timos Sellis during the third FLaReNet
Forum, a clear understanding is needed of the pros and cons of closing or opening up data:
“What do we gain by closing data? Are business models based on closed and heavily guarded LR
actually successful? Are we losing opportunities for growth by not systematically exploiting
common sharing and synergies? What do we lose by opening up data? Can the direct income lost
be compensated by direct or indirect financial gains?” (Sellis 2011).
To share resources, both data and tools, has become a viable solution towards encouraging open
data, and the community is strongly investing in facilities for the discovery and use of resources
by federated members. These facilities, such as the META-SHARE infrastructure, could represent
26
Ch 1. The Blueprint
an optimal intermediate solution to respond to the need for data variety, ease of retrieval, better
data description and community-wide access, while at the same time assisting in clearing the
intricate issues associated with IPR (see also DiPersio, 2011).
A KPMG study on the Canadian Spatial Data Infrastructure (Sears, 2001), concluded that closed,
restricted data has major economic harm: "the consequences [of cost recovery] for businesses
are higher marginal costs, lower research and development investments and threatened
marginal products. The results for consumers are negative: higher prices and reduced products
and services. The overall economic consequences... are fewer jobs, reduced economic output by
almost $2.6 billion and a lower gross domestic product." One branch of Language Technology,
Question Answering, has benefited greatly from the availability of open data resources,
especially the research datasets created for the yearly TREC question answering track. Thinking
along these lines, the availability of massive quantities of open data could transform the NLP
industry, as suggested by J. van der Meer for translation (van der Meer 2011). The success of the
sharing approach is well represented by TAUS 13, where many translation leaders have started
sharing their translations in the Taus Data Association repository.
A questionnaire carried out by FLaReNet strongly advised that at least those resources that were
developed with public funding should be made openly available. Another suggestion is to ensure
openness of resources for most types of uses, making use of standardised licenses where
available,
Of course, certain types of data are and will probably remain not shareable, either for
confidentiality reasons, personal data, or competitive ones. Non shared data can still be
exploited, for instance through shared services (see also Ch. 4 and 7).
In the meantime, it is important to define appropriate criteria for different levels of openness
and to define “best practices” for making resources available that address different constraints
that may be faced.
However, achieving large-scale, open datasets is only one of many general requirements for
rapid, open advancement of language technologies. As pointed out during the last FLaReNet
Forum, it only makes sense to talk about data openness after clearing the discovery and
reusability stumbling blocks. Before data can be openly usable, they need to be a) easily
retrievable and b) easily reusable. Point a) is addressed by the Documentation issue in 1.3, while
reusability has to do with resource Interoperability (see 1.5). Whether making data open or not
is a matter of choice at the licensing level, it crucially depends on choices made at the very early
stage of resource planning and production, i.e. by ensuring content and formal interoperability
and proper documentation.
Legal, IPR issues
IPR issues are crucial to facilitating growth in our sector, yet they pose real problems. On the one
hand IPRs (especially authorship) need to be protected; on the other IPR tends to restrict
accessibility to and usability of language resources.
We do not yet have a sufficient grasp of the trans-border legal issues in the EU to support
enhanced resource sharing and legally protect LRs against improper reuse, copying,
modification etc. The Berne Convention for the Protection of Library and Artistic Works extends
copyright protection to creators in countries other than their own, but enforcement is still a
national issue and is therefore implemented in different ways.
On top of this, the availability and use of huge quantities of web data as useful resources creates
a novel situation that raises further legal problems. Legislation is lagging behind the technology.
The current trend is towards a culture of free/open use with less protective holders’ rights.
Creative Commons, for example, is one of the most widely used license models for language
resources (see Google, Wikipedia, Whitehouse.gov, Public Library of Science, and Flickr).
13
http:// www.translationautomation.com/
27
Ch 1. The Blueprint
The LR community is also facing major questions about how to use blogs, newsgroups, web
video, SMS and social network sites as data, as there are virtually no laws, regulations or court
decisions governing these media. This means that most resources are kept in-house for research
or use, otherwise individuals and organisations risk law suits due to some form of infringement.
The challenge for both the LRT community and policy makers is to push for the development of a
common legal framework that would facilitate resource sharing efforts that do not break the law.
It is crucial to disseminate a certain amount of legal knowledge/know-how to educate all
(major) players in the LRT area. It is also important to inform a number of lawyers about
community concerns so they can develop adequate frameworks to address such issues.
Moreover, it is important that such legal experts are asked to intervene in the initial phases of
resource production, to ensure that all legal (and also ethical, privacy and other) aspects are
taken into consideration when planning for long-term LR sharing and distribution.
The community should also avoid one-size-fits-all solutions. There are a large number of
licensing schemes already in use today, some are backed by strong players (ELRA, LDC, open
source communities such as Creative Commons, GPL, etc.), others have been drafted bilaterally
and in some cases by the legal departments of data providers. It is crucial that such licensing is
harmonised and even standardised. Licensing schemes need to be simplified through broadbased solutions for both R&D and industry. Electronic licensing (e-licenses) should be adopted
and current distribution models to new media (web, mobile devices, etc.) should be accepted.
For mixed-funded initiatives (private/public), we should ensure that there is an agreement to
make resources available at fair market conditions right from the start.
1.8
Resource Sustainability
“Ensure sustainability of language resources”
LRP
PM
Ensure that all LRs to be produced undergo a sustainability analysis as part of the
specification phase
•
•
Foster use of a sustainability model
•
Make sustainability plans mandatory in projects concerning LR production
Assess LR sustainability on the basis of impact factors
•
•
•
•
Sustainability covers preservation, accessibility, and operability (among other things) that all
have mutual influences. Currently, most resource (data and software) building and distribution
is based on short-term projects, which often leads to the loss of resources when the projects end.
Collecting and preserving knowledge in the form of existing LRs instead should be a key priority.
LRs must be accessible over the long term. This means:
− Archiving and preserving the data by the production unit, and also archiving them off-site
(e.g. in very-long term archiving/data centres).
− Maintaining LRs is an appropriate way.
−
28
Making sure that linguistic tools and resources are sustainable, e.g. by requesting resource
accessibility and usability for a given time frame.
Ch 1. The Blueprint
The challenge here is to establish clear rules for properly archiving and preserving resources
without making resource creation and sharing process too complicated.
A top priority is developing an analytic model of sustainability in which extrinsic and intrinsic
factors are taken into account, together with new methods and practices for sharing and
collaboration 14. The model must be backed by appropriate initiatives to encourage
implementation.
1.9
Resource Recognition
“Promote the LR ecosystem”
LRP
Develop a standard protocol for citing language resources
Carry out a rolling survey of conference papers and LRs mentioned in papers.
Identify the “weak signals” indicating new trends in LT and LRs internationally
Bring together the main LR stakeholders to find a way of attach a Persistent and
Unique Identifier (PUId) to LR
•
Give greater recognition to successful LRs and their producers (Prizes, Seals of
Recognition, etc. )
•
Support training in production and use of LRs
•
Track the use of LRs in scientific and professional papers
Compute a Language Resource Impact Factor
Introduce training in the production and use of LR in Computational Linguistics
and Language Technology curricula
Continue to organize tutorials on LR production at conferences such as LREC
•
•
•
PM
•
•
•
•
•
•
Language Resources (both data and software) are time-consuming, costly and increasingly
require a considerable share of research budgets. The entire ecosystem around Language
Resources needs substantial support and recognition. Small labs and individual researchers are
not keen on depositing or sharing their resources because there has been little incentive to do
so. There are very almost no rewards for researchers and institutions to share, preserve and
maintain resources, and this now poses a number of serious problems.
Language Resources thus deserve credit and should be cited in a similar way to sources in
scientific publications. A model for citing LRs would therefore be highly desirable such as a
standard citation framework that would allow for citing LRs in a uniform way (this would also
enforce the use of minimal metadata descriptions) and for which LR providers will be
responsible and credited for.
Along the lines followed in other fields, especially in Biology, a “Language Resources Impact
Factor (LRIF)” should be defined in order to enforce the practice of citation of resources on the
model of scientific paper authoring and to calculate actual research impact of resources.
14
See the model proposed by FLaReNet in Ch. 2.
29
1.10
Ch 1. The Blueprint
International Cooperation
“Promote synergies among initiatives at international level”
LRP
Share the effort for the production of LRs between international bodies, such as
the EC for the EU, and individual countries/regions, such as Member States in
Europe
Encourage agencies to open their calls to foreign participants
Establish an International Forum to share information, discuss strategies and
declare/define common objectives.
Establish MoU among existing initiatives, starting with the EC efforts
•
•
Maintain a public Survey on the LT and LR situation worldwide, based on
FLaReNet and META-NET
•
Produce a White Paper on LT, including LRs, in preparation for FP8. Distribute it
within Europe and abroad
•
Ensure the sustainability of the community driven initiatives such as the LRE
Map, META-NET Language Matrixes, and FLaReNet Network of International
Contact Points and National Initiatives Survey Wiki
PM
•
•
•
•
Cooperation among countries and programs is essential (particularly for infrastructure) to drive
the field forward in a coordinated way and avoid duplication of efforts and fragmentation.
A coordinated effort at the international level would help by providing less advanced
countries/languages with examples and best practices, such as defining a commonly agreed on
set of basic LRs that have already proven necessary for producing LTs efficiently for better
represented languages. This kind of international effort should also try to identify the gaps and
draw up an appropriate roadmap to fill them.
International cooperation between infrastructure initiatives is also important for avoiding the
duplication of effort, ensuring that standards are truly international, and encouraging the free
exchange of ideas.
It is crucial to discuss future policies and priorities for the field of Language Resources and
Technologies not only on the European scene, but in a worldwide context. This is true both when
we try to highlight future directions of research, and – even more – when we analyse which
infrastructural actions are needed. The growth of the field must be complemented by a common
effort that looks for synergies and overcomes fragmentation.
It is an achievement, and an opportunity for our field, that recently a number of strategicinfrastructural initiatives have started, or are going to start, all over the world. This is also a sign
that funding agencies recognise the strategic value of the LR field and the importance of helping
a coherent growth also through a number of coordinated actions.
Cooperation is an issue that needs to be prepared. FLaReNet has been the place where these
initiatives get together to discuss and promote collaboration actions.
Networking and support actions must be conducted more intensively, with establishment of
international committees that have formal recognition. In a field that is both fragmented and
over-structured, many mentioned the need to have an International Forum (a meta-body) to
share information, discuss strategies and declare/define common objectives. Such a Forum can
play a role only if it is recognised as influential and authoritative: e.g. a Memorandum of
Understanding signed by hundreds of organisations could give authority.
30
Ch 1. The Blueprint
References
Chris Callison-Burch, Mark Dredze. 2010. “Creating Speech and Language Data With Amazon’s
Mechanical Turk”. In Proceedings NAACL-2010 Workshop on Creating Speech and Language Data
With Amazon’s Mechanical Turk.
Nicoletta Calzolari, Monica Monachini, Valeria Quochi, Núria Bel, Gerhard Budin, Tommaso
Caselli, Khalid Choukri, Gil Francopoulo, Erhard Hinrichs, Steven Krauwer, Lothar Lemnitzer,
Joseph Mariani, Jan Odijk, Stelios Piperidis, Adam Przepiórkowski, Laurent Romary, Helmut
Schmidt, Hans Uszkoreit, Peter Wittenburg. 2011. The Standards’ Landscape Towards an
Interoperability Framework. The FLaReNet proposal. FLaReNet 2011.
Kenneth Church. 2011. “Plan B”. In Proceedings of the Third FLaReNet Forum, Venice, Italy, 26-27
May 2011.
Denise DiPersio. 2011. “Is our relationship with open data sustainable?”. In Proceedings of the
Third FLaReNet Forum, Venice, Italy, 26-27 May 2011.
Bill Dolan. 2011. “Parallel Multilingual Data from Monolingual Speakers”. In Proceedings of the
Third FLaReNet Forum, Venice, Italy, 26-27 May 2011.
David Giaretta. 2011. “Preparing to share the effort of preservation using a new EU preservation
e-Infrastructure”. In Proceedings of the Third FLaReNet Forum, Venice, Italy, 26-27 May 2011.
High Level Expert Group on Scientific data. 2010. “Riding the Wave: How Europe can gain from
the rising tide of scientific data”. Final report of the High level Expert Group on Scientific Data.
October 2010.
Nancy Ide, James Pustejovsky. 2011. “An Interoperability Challenge for the NLP Community”. In
Proceedings of the Third FLaReNet Forum, Venice, Italy, 26-27 May 2011.
Jaap van der Meer. 2011. “Imagine we have 100 Billion Translated Words at our Disposal”. In
Proceedings of the Third FLaReNet Forum, Venice, Italy, 26-27 May 2011.
Dunja Mladenic, Marko Grobelnik. 2011. “Cross-lingual knowledge extraction”. In Proceedings of
the Third FLaReNet Forum, Venice, Italy, 26-27 May 2011.
Daniel Pimienta, Daniel Prado and Álvaro Blanco. 2009. Twelve years of measuring linguistic
diversity in the Internet: balance and perspectives. UNESCO, Paris.
Frederique Segond. 2011. “Turning water into wine : transforming data sources to satisfy the
thirst of the knowledge era”. In Proceedings of the Third FLaReNet Forum, Venice, Italy, 26-27
May 2011.
Garry Sears. 2001. “KPMG Consulting Inc. for GeoConnections Policy Advisory Node”, Canadian
Geospatial Data Policy Study, March 2001.
Timos Sellis. 2011. “Open Data and Language Resources”. In Proceedings of the Third FLaReNet
Forum, Venice, Italy, 26-27 May 2011.
Andrejs Vasiljevs. 2011. “How to get more data for under-resourced languages and domains?”. In
Proceedings of the Third FLaReNet Forum, Venice, Italy, 26-27 May 2011.
Omar Zaidan, Chris Callison-Burch. 2011. “Crowdsourcing Translation: Professional Quality from
Non-Professionals”. In Proceedings of ACL-2011.
31
Ch 2. Sustainable LRs
Chapter 2 - Identifying Mature and Sustainable
Language Resources
Khalid Choukri, Joseph Mariani
This chapter looks at LR sustainability and maturity on the basis of a descriptive model of
sustainability from a LT perspective. It summarizes the description of a sustainability model,
elaborated within the project 1.
The sustainability as widely understood in the HLT field is the ability of a given Language
Resource to survive over time without any explicit and external financial support, which could
be referred to as a self-sustained resource. When mentioning such an expression, no reference is
made to the availability and use by the wider community (whether R&D or industry), no
reference is made to its rights management (e.g. licensing), to the ability to be customized to suit
new needs, to its “openness” so new users could reshape it and convert it to an operable
resource in their environment, no reference is made to its updates, corrections, improvements,
repackaging, etc. These are however crucial parameters and the proposed model intends to list
an exhaustive number of factors that impact this sustainability. These factors have also been
clustered into Pillars (or categories) to ensure better understanding and ease to use in the
prediction of sustainability of LRs.
In addition to the "sustainability" term introduced herein, we often encounter other terms such
as preservation, perpetuation, etc.
The goal of such descriptive model is to supply LR producers, packagers, maintainers, datacentres/distributors, users as well as funding agencies with appropriate tools that should help
them design a rational and cost-effective lifecycle of LRs. Such model will be hinted at and will be
implemented as risk-management-model, that could be used as a tool to anticipate on factors
that may not comply with the sustainability requirements.
Maturity can be defined with respect to the following LT pyramid and the corresponding
building blocks as shown in the following diagram:
1
see FLaReNet Deliverable D2.2a for details.
32
Ch 2. Sustainable LRs
Involvement
of early
Adopters
Maturity of "Market"
(Market demand) and
Adopters
Maturity of applications
Maturity of research, transfer of technologies
Availability of existing LRs with a sharable, "critical mass"
Knowledge of LR specifications, production, quality assessment, etc.
The most important level is the lowest – knowledge of LR specifications, production, and quality
assessment. It is fairly easy to assess the maturity of an area of research or application involving
LRs when there is easily available public knowledge and know-how about the specifications
(technical, logistic, linguistic, legal, etc.). Some of this knowledge may even come from
established technical standards.
A framework such as SPEECHDAT 2 is an excellent example of resources developed in a welldesigned configuration and broadly shared within the speech recognition community. The
approaches adopted to develop such resources (adapted to R&D and pre-competitive tasks)
have been widely used to produce resources that can be used in real applications. For a large
number of technologies, there are well established best-practices (e.g. for SMT and broadcast
news speech to text.).
2.1
PM
Ensure strong public and community support to production and dissemination of “bestpractices”
The second most important level in the pyramid is the availability of a critical mass of LRs. Mere
knowledge about LR production best practices will not boost R&D unless these resources are
developed in an adequate framework. This could lead to sharing resources and comparing the
outcome of research activities between competing-cooperating labs. This capacity can enhance
the maturity of the research and facilitate any subsequent technology transfer. The existence of a
critical mass of LRs can trigger two contrasting types of initiatives:
2
More details on www.speechdat.org
33
Ch 2. Sustainable LRs
•
Boosting public and/or community support for sharing existing LRs of paramount
importance for the EU, or
• Leaving it to the market (or other non EU public bodies).
A third level is the maturity of research and of the research community in question for a given
topic. It is important to know whether there is a critical mass of research/researchers focused
on a particular technology. This requires a careful monitoring of innovative work, scientific
publications, patent applications, etc.
If it is unlikely that critical mass will be reached in a given area, policy makers will either deem it
a vital/sensitive area (e.g. cultural heritage or security) and launch strong supporting programs,
or it will be left to the market, where there is a risk that non-European economies could take the
lead (as happened in the Personal Computer industry in the early 1990s).
2.2
LRP
2.3
PM
Monitor on-going research developments through scrutinizing publications and patent
filings, and keeping track of prize-winning papers and Theses
Support EU-critical LT domains once they have been identified. MT, Speech to Text
(transcriptions), and Human Machine Interactions, are obvious examples
Research maturity can be assessed objectively through evaluation campaigns built on a
framework of test sets, metrics, and technology providers (see also chapter 5 on evaluation). It
can also be assessed by analysing R&D-to-Market technology transfer.
2.4
2.5
LRP
PM
PM
Ensure that there are regular evaluation campaigns to assess the state of the art and
provide an accurate picture of EU research and technology
Conduct regular surveys of technology transfer (ideally to stimulate, not simply reflect
it)
The fourth level refers to how far a tool, prototype, or component can be used virtually "out-ofthe-box" (i.e. with an acceptable level of engineering effort).
Finally market maturity. Market demand and the existence of (early) adopters cannot be
guaranteed when there is a full-equipped set of available technology and resources in place.
Imperfect technologies can create highly successful applications, and brilliant technologies can
fail in the marketplace due to lack of take-up or relevance.
2.1 Key sustainability factors
Reviewing the life cycle of several LRs, we identified a large number of extrinsic and intrinsic
features that have an impact on the life of these resources. We have tried to be as exhaustive as
possible while avoiding overlaps and redundancies. Their impacts on the life of resources have
34
Ch 2. Sustainable LRs
been given estimated weights based on our experience and educated guesses. The use of these
factors and their respective weights led to our sustainability score.
The factors that have been defined3 as those impacting LR sustainability are listed herein:
1) Specifications (including references to best-practices & standards)
2) Production and management of the documentation
3) Quality assessment and/or quality validation report
4) Management of rights, ethics, privacy, consent, and other sensitive legal issues (before
and during production)
5) Information dissemination including scientific publications
6) Formats, encoding, content
7) Portability across languages, environments and domains
8) Packaging (a compilation of all parts including resource documentation)
9) Management of use rights and licenses for sharing and distribution
10) Data identification, metadata and discovery
11) Versioning and referencing
12) Usability assessment and relevance
13) Accessibility (in packages, over the web or other)
14) Accessibility of LRs in an “open” mode
15) Preservation of media to ensure long-term access
16) Access charge (for free /for a fee)
17) Reference to production and projects, environments in which such resources have been
produced/used
18) Relevance for other NLP applications and areas
19) Maintenance and support over time
20) Role and impact of large-scale data centres and archiving facilities.
In order to stress the important "categories" of impacting factors, these have been clustered into
8 groups:
a) Proper documentation with appropriate documentation management rules
b) Management of rights, ethics, privacy, consent, and other sensitive legal issues
c) Information dissemination including scientific publications
d) Rights and licenses
e) Data identification, metadata and discovery
f) Accessibility of packages and media
g) Accessibility of LRs in “open” mode
h) Relevance to more advanced NLP applications and areas
Note that extrinsic factors such as documentation, licensing, and access are among the key
criteria for making a resource usable in the wider community. For example, despite theoretical
controversies about the nature of the original WordNet, researchers and other users have
developed and actively supported a similar resource for their own language (over 70 languages
now have WorldNet related projects) largely because the documentation, rights, and access
were properly addressed and maintained, making it easier for a critical mass of research and
researchers to feel confident about using them.
Example of the scoring diagram could be:
3
See FLaReNet deliverable D2.2a, “Identification of mature self-sustainable LRs versus areas to be sustained and
sustainability factors”.
35
Ch 2. Sustainable LRs
2.6
2.7
LRP
LRP
PM
Develop a sustainability scoring tool and make it publicly available4
Assess LR sustainability on the basis of the sustainability factors and ensure that all LRs
to be produced undergo a sustainability analysis as part of the specification phase
2.2 Key areas of research interest
Key areas of interest are identified on the basis of the technologies that have been discussed
within various forums over the last three years. The domain’s own taxonomy has been under
discussion for the last decade and there is still no consensus about the best way of cutting up
this complex cake. We can look at two obvious dimensions:
In terms of media (also called modality):
• Text (corpora annotated at various levels: POS, syntax, semantics, coreference/anaphora; lexica, thesauri, ontologies; translation memories/aligned corpora,
etc.)
• Speech, audio, and visual (including multimedia) (transcriptions, signal/phonetic
alignments, etc.)
• Video (scene annotations, speaker identification, topic detection, etc.)
• Other (sign languages, etc.)
Or application type:
• Web search, retrieval, extraction, indexing, named entity recognition, opinion mining, QA answering, multimedia, cross-lingual IR
• Text and discourse analysis & mining, summarisation, plagiarism detection
4
See a preliminary version at http://www.elda.org/flarenet_sustainability_dimensions.html
36
Ch 2. Sustainable LRs
•
•
•
•
2.8
Machine translation, including speech to speech translation
Speech recognition, speaker recognition, identification (voice command and Interactive
Response Systems, dictation systems, conversational speech, speech to text from
broadcasters, etc.), Speech synthesis (text-to-speech, avatars, etc.)
Dialogue management (speech and/or text based)
Language generation (sentence, report, etc.) associated with emotions, speech synthesis,
etc.
PM
Maturity has to be discussed and elaborated upon according to the two “description”
dimensions (modality, application)
2.3 LR maturity and sustainability in major research areas
National Text Corpora
One key research stream has been the collection and analysis of various features of corpora of
written and spoken data. A quick survey of existing corpora shows that most East-European
languages have at least one major reference corpus, often referred to as a National Corpus. This
is the case for Bulgarian 5, Czech 6, Hungarian7, Polish 8, Slovene 9, Slovak 10, etc. Similar corpora
already exist for UK English (the famous British National Corpus 11) and now American English
(American National Corpus and in particular the freely available sub-set - Open ANC or OANC 12).
Many of these "National" corpora reflect the current use of the general language and are
appropriate for LT purposes, though for a number of them the content is largely skewed to the
Social Sciences and Humanities (incorporation of sources for which copyright "period" expired).
The major drawback of such resources is they are often copyrighted by third parties and hence
are only offered through access tools (via an online service with sophisticated query forms that
extract "samples" to better understand the language features of this domain of knowledge). Such
approach does not allow exploiting the whole corpus, which is essential for developing LT
applications.
Corpora specifically produced for LT tend to gain some legitimacy from the clout of the hosting
institute, but also from the wealth of levels of annotation and mark-up (POS, syntax, coreferences, anaphora, and other types of annotations).
 Maturity: Corpora used in LT (e.g. the BNC) have benefited from over a decade of
feedback, tuning, and corrections. But many others are used as online services and are of
little use for LT (given the access mode that limits their availability).
 Sustainability: High because of the institute or consortia underwriting the resource.
2.9
5
6
7
8
9
LRP
Ensure that efforts are put in clearing the IPR issues and reformatting the data, so that
http://ibl.bas.bg/en/BGNC_en.htm
http://ucnk.ff.cuni.cz/english/index.php
http://mnsz.nytud.hu/index_eng.html
http://www.nkjp.pl/index.php?page=0&lang=1
http://www.fidaplus.net/Info/Info_index_eng.html
http://korpus.juls.savba.sk/index_en.html
10
11
12
http://www.natcorp.ox.ac.uk/
http://www.americannationalcorpus.org/
37
Ch 2. Sustainable LRs
such (national) corpora could be repurposed and repackaged for LT
Spoken Corpora
Speech corpora used in LT have been developed to adapt to the evolution of research topics,
mostly influenced by market and security concerns 13.
The first R&D programs targeted command and control (small vocabularies), followed by
dictation, interactive tasks (e.g. ATIS and Sundial 14), and then to transcription of broadcast news
and conversational speech. The main focus today in speech technology is on the transcription of
audio speech (broadcast news, meetings, lectures, conversational speech, etc.), information
retrieval from audio/transcribed data, speech-to-speech translation, the automatic
summarization of spoken data, etc.
There are corpora of transcribed broadcast news for a number of languages (American/British
English, Arabic, French, German, Portuguese, Thai, Vietnamese, etc.) but the research community
does not always have easy access to them. Major evaluation campaigns (e.g. NIST, ELDA) focus
on a few languages (e.g. English, Mandarin Chinese, French and Arabic) so more use is made of
English corpora, for example, than any other language.
A few years ago, the telephone-based speech recognition community began a large programme
on front-end evaluations called Aurora 15, led by the Aurora Working Group at the European
Telecommunications Standards Institute (ETSI). The purpose was "to determine the robustness
of different front ends for use in client/server type telecommunications applications." To
support this initiative, a number of the resources developed in this programme have been repackaged, and Aurora databases are available from ELRA.
Other resources for Human-Machine Interactions have been produced but cover artificial
settings, as none of the major users of dialogue systems and/or human-human services (such as
call-centre operators) are willing to share their data with the R&D community and potential
competitors (including reasons of privacy). The "artificial" setting is often based on Wizard of Oz
simulations (MEDIA and Port-Media projects at ELDA 16) or conversations between friends and
relatives on pre-defined topics (e.g. Call Friends & Call Home at LDC 17).
18
 Maturity can only be reached when a critical-mass of community players decides to
tackle the same issue such as broadcast news, Aurora front-ends, etc. There are still
major technical problems with conversational speech and broadcast news collected in
low-bandwidth contexts, noisy channels, overlapping speech in intense discussion
contexts, etc.
 Sustainability: Resources from responsible data centres are usually sustainable (i.e.
they can be identified and discovered by newcomers), but many resources developed
and used internally by individual organizations cannot be accessed and maintained.
Corpora for MT
During the last decade, work on Machine Translation has been fostered by the statistical
approach using translation memories and aligned corpora as training and tuning resources.
Aligned corpora have stimulated the development of MT systems for pairs of languages for
13
14
15
16
17
18
38
See performance illustration in section 2.7 as given by NIST.
The ATIS resources are available through LDC while the EC project Sundial resources are lost!!
http://www.elda.org/rubrique18.html
http://www.elda.org/article252.html
www.ldc.upenn.edu/Catalog/LDC96S49.html
Or more often a funding agency!
Ch 2. Sustainable LRs
which traditional rule-based MT systems would have been too expensive to craft. Most of these
resources are available either for the major languages used in international bodies (EU and UN),
or for those that are used by a large software development community (on the basis of produced
technical manuals).
Most EU languages, for example, are well supplied with resources from the Europarl and JRCAcquis corpora. This is also true for languages used in large international organizations e.g.
Arabic, Russian, Chinese, French, English and Spanish at the United Nations, part of whose
multilingual holdings have been released. This has enabled baseline MT systems to be built at a
low cost by R&D teams, many of which use MOSES open source software (see for instance the
baselines developed for English to Arabic within the Medar project 19).
For other languages, open resources have been packaged, exploiting translated and localised
software technical manuals and screen messages e.g. Linux manuals, KDE manuals (KDE Kool
Desktop Environment manuals), etc.
One side-effect of this statistically-driven, aligned corpus approach is the development of more
research into crawling techniques and tools as well as alignment tools. Exploitation of alignment
tools for the extraction of bilingual lexica has been carried out by many labs that have improved
such tools.
Emerging hybrid approaches to MT combine statistical with rule-based techniques. Some SMT
engines mix together aligned “raw” corpora with aligned POS tagged corpora to make use of
more knowledge and leverage existing resources.
The evaluation of MT systems in large campaigns carried out by NIST (annual MT evaluation),
ELDA (TC-STAR, CESTA and MEDAR), Euromatrix/EuromatrixPlus projects, have also helped
underwrite the SMT approach for a number of language pairs. Existing resources are not enough
and more effort should be put into compiling adequate resources in collaborative environments.
 Maturity: In SMT, (relatively few) fairly small-scale existing resources have been used
and reused again, but there are also some larger-scale resources for some EU and UN
languages.
 Sustainability: Many of the language resources involved are supported by strong
institutions (the EC Joint-Research Centre in Ispra, ELRA, LDC, NIST, etc.) and are
therefore likely to survive.
2.10
LRP
2.11
LRP
Implement tools that can help crawl “comparable” resources, select a small set that can
be aligned and then provide clear legal conditions to ensure they can be shared with the
MT community and beyond (assuming these can be annotated for bilingual lexical
extraction, named entities, etc.)
Develop a certain volume of MT-based corpora, post-edit them (semi)automatically to
specific quality levels, and then use them iteratively to train new MT systems more
effectively
2.4 Evaluation packages and resources
A large number of technology areas are backed by well-designed and well-organised evaluation
campaigns, DARPA being an exemplary case in point with sustained work on MT, audio
transcription, and speaker identification and other topics that have been evaluated by NIST (and
by LDC for the resources). The EC and European national governments have also backed
19
http://www.medar.info/MEDAR_evaluation_package/package_medar_v5.tar.gz
39
Ch 2. Sustainable LRs
evaluation efforts such as Technolangue and Technovision in France (ELDA) and N-best/STEVIN
in the Netherlands, while Evalita in Italy has been a voluntary community effort.
This approach highlights mature practices for producing language resources, and assesses the
maturity of specific technologies. To ensure comparability of the results, participants have to
agree on a common framework for data content as well as formatting and encoding
specifications.
 Maturity: Evaluation campaigns are clearly the best way to assess the maturity of a
given type of technology/resource; evaluation centres need to agree on a number of
features, and such agreement (consensus) only become feasible if topics have reached a
certain level of maturity.
 Sustainability: Possible when an official body underwrites the availability of the whole
evaluation package and resources, but difficult when it comes down to project level.
2.12
LRP
Define a small set of crucial topics, appropriate for comparable evaluations, on which a
critical mass of research/researchers should focus through a clear adherence to
principles
The ELRA HLT evaluation portal 20 compiles information about past and present LT evaluation
activities. It has identified a large number of multilingual and monolingual areas where activities
have been carried out with specific resources, ensuring that best-practices for developing test
data have been agreed upon by those involved.
For instance the ImageCLEF 21 evaluation (in its 9th year) has developed methods for releasing
test data in areas such as:
• Medical retrieval
• Photo annotation
• Plant identification
• Wikipedia retrieval
• Patent image retrieval and classification in collaboration with CLEF-IP
The same goes for NIST that assesses "progress in content-based analysis of and retrieval from
digital video via open, metrics-based evaluation" in its video retrieval evaluation (TRECVID 22).
Areas identified include:
•
•
•
•
•
•
•
•
•
20
21
22
40
Machine Translation
Speech-to-Speech Translation
Multilingual Text Alignment
Information Extraction
Information Retrieval
Text Summarization
Parsing
Speech Recognition
Speech Synthesis
http://www.hlt-evaluation.org/
http://imageclef.org/2011
http://www-nlpir.nist.gov/projects/tv2010/tv2010.html
Ch 2. Sustainable LRs
• Multimodal Technologies
These multimodal technologies include technologies combining features extracted from
different media (text, audio, image, etc.):
• Audio-visual Speech Recognition
• Audio-visual Person Identification
• Audio-visual Event Detection
• Audio-visual Object or Person Tracking
• Biometric Identification (using face, voice, fingerprints, iris, etc.)
• Head Pose Estimation
• Gesture Recognition
• Multimodal Information Retrieval (e.g. Video Retrieval)
• etc.
Related resources have been developed in projects such as AMIDA, QUAERO, TRECVID,
IMAGECLEF, CHIL, VACE, TECHNO-VISION, BIOSECURE, etc.23
On the basis of such lists, we can conclude that many of these areas are mature enough in terms
of best practices to be able to produce resources for training and/or benchmarking the
technologies. Most of these, however, only cover a small set of languages or language pairs. For
instance Summarization data can be found in projects such as Must/NTCIR (Japan), TAC/NITS,
TIPSTER, TIDES (USA), GERAF (Canada, French) that in all only cover three or so languages.
2.13
PM
Ensure that resources are developed for all EU languages so that evaluations can be
made of language-dependent technologies listed by the major agencies (training and
testing data)
2.5 Multiple-application resources as a "maturity/sustainability"
factor
Some LRs can be used to develop and test technologies in more than one LT research or
technology area. This is important for sustainability as it enables resource production and
maintenance costs to be shared across larger communities over longer spans of time.
A good example is the set of large monolingual corpora that are annotated at various analytic
levels - morphological, named entity, WSD, and treebank - and are often used by the speech
community to build language models, and by the text community to test named entity
recognition, syntactic analysis, and so on. The main development problem is not being able to
anticipate the appropriate range of coding/encoding features that will ensure that a given
resource can be used effectively for new NLP applications as they emerge.
Generally speaking, if a resource is backed by a large community of evaluators (which implicitly
indicates adherence of a large number of researchers and developers) , there is better chance
that it will be more widely used.
2.6 Maturity as reflected by the LRE MAP (2010) data
The LRE Map 24 shows the "big" trends, but is not granular and synthetic enough to identify how
mature each resource is for each language.
23
24
http://www.hlt-evaluation.org/spip.php?article149
http://www.resourcebook.eu
41
Ch 2. Sustainable LRs
Bulgarian
Czech
Danish
Dutch
English
Estonian
Finnish
French
German
Greek
Hungarian
Irish
Italian
Latvian
Lithuanian
Maltese
Polish
Portuguese
Romanian
Slovak
Slovene
Spanish
Swedish
Other Europe
Regional Europe
Multilingual
L.I.
N.A.
Total
The Language Matrixes produced in META-NET on the basis of the LRE Map data (Mariani and
Francopoulo 2011) provide a synthetic view for the various modalities (Written language,
Spoken language, Multimodal/Multimedia) and for different categories of Language Resources
(Data, Tools, Evaluation means and Meta-resources (Standards, Metadata, Guidelines), but they
still lack granularity and completeness. The Language Matrix on Multimodal/Multimedia data
below shows for example that there is a strong prominence of the English language, with 22 of
the 82 identified corpora 25. But it could be interesting to have a more detailed analysis of the
different types of corpora, while the situation for the other languages and for the other types of
resources is sparser and needs an improved taxonomy for the Types (the initially suggested
ones are in bold characters) and more data from new sources.
MM Data
(Ranked)
Corpus
Terminology
Database
Ontology
A complete archive of resources
Data Collection and Annotation
Management System
Grammar/Language Model
Lexicon
Total
0
0
0
0
0
1
0
0
0
0
3
0
0
0
0
4
0
0
0
0
22
2
1
0
0
1
0
0
0
0
2
0
0
0
0
3
1
0
0
0
10
2
0
0
0
2
0
0
0
0
2
0
0
0
0
0
0
0
0
0
4
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
4
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
7
1
0
0
0
4
0
0
0
0
3
0
0
0
0
1
0
0
0
0
0
0
0
0
1
3
1
0
2
0
5
0
1
0
0
82
7
2
2
1
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 1 3 4 25 1 2 4 12 2 2 0 4 0 1 0 0 4 0 0 0 8 4 3 1 1 7 7 96
E.1.a. Multimodal/Multimedia Data (Ranked order)
The same goes for written language data. Over 950 resources have been identified, the top three
covering English with 327 items, French with 80 items, and German with 70. More than 50% of
those resources (552) are constituted by corpora, and 25% by lexica (229) but there is a long tail
of new Types entered by the authors.
25
42
The types of resources that appear in bold characters correspond to the types suggested in the questionnaire. The
other ones have been added by the authors.
7
Corpus
6
Lexicon
1
Ontology
1
Grammar/Language Model
1
Terminology
A syntactic judgments database 0
Resource: morphology
0
Thesaurus
0
A Knowledge Base with
0
Lexical-Semantic Relations
between words
A list of categories with
0
examples of language use
Controlled Legal Language
0
Corpus-Based Online
0
Dictionary
Database
0
Encyclopedic knowledge
0
Event Semantics
0
Lexicon/corpus
0
Online Encyclopedia
0
Repository of bilingual
0
lexicons
Resources integration
0
Text Book
0
Virtual Game World
0
Yahoo!'s local listings in
0
Chicago
Encyclopedia
0
Part-of-Speech Tagset
0
Psycholinguistic Database
0
Total
16
12
7
2
1
1
0
0
0
6
2
0
2
0
1
0
0
17 206
8 77
2 18
1 11
2 10
0 0
0 1
0 1
3
1
0
0
1
0
0
0
3
2
0
1
0
0
0
0
44
24
3
4
5
0
0
0
43
15
4
2
3
0
0
1
10
3
2
0
0
0
0
0
8
4
0
1
1
0
0
0
1
0
0
0
0
0
0
0
32
16
4
2
0
0
0
0
9
0
0
0
1
0
0
0
4
0
2
0
1
0
0
0
1
0
0
1
0
0
0
0
7
2
1
2
1
0
0
0
19
6
1
1
0
0
0
0
12
7
1
1
0
0
0
0
2
0
0
1
0
0
0
0
5
1
0
0
0
0
0
0
29
19
4
5
2
0
1
0
19
4
0
1
0
1
0
0
19
11
3
3
2
1
0
0
18
8
0
1
3
0
0
0
5
3
1
0
1
0
0
0
9
3
16
2
1
0
0
0
Total
Estonian
Finnish
French
German
Greek
Hungarian
Irish
Italian
Latvian
Lithuanian
Maltese
Polish
Portuguese
Romanian
Slovak
Slovene
Spanish
Swedish
Other Europe
Regional Europe
Multilingual
L.I.
N.A.
English
Written Data
(Ranked)
Bulgarian
Czech
Danish
Dutch
Ch 2. Sustainable LRs
2 552
0 229
2 67
1 45
0 36
0 3
0 2
0 2
0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0
1
0 0 0
1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1
0 0 0
0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1
0 0 0
0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
1
1
0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0
1
0 0 0
0 0 0
0 0 0
1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0
1
1
1
0 0 0
1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1
0
0
0
23
0
0
0
0
0
0
0
0
11
0
0
0
0
0
0 0
0 0
0 0
30 327
0
0
0
0
0
0
0
0
5
0
0
0
0
0
0
0
0
6
0
0
0
0
0
0
0
0
80
0
0
0
0
0
0
0
0
70
0
0
0
0
0
0
0
0
15
0
0
0
0
0
0
0
0
14
0
0
0
0
0
0
0
0
1
0
0
0
1
1
0
0
0
56
0
0
0
0
0
0
0
0
10
0
0
0
0
0
0
0
0
7
0
0
0
0
0
0
0
0
2
0
0
0
0
0
0
0
0
13
0
0
0
0
0
0
0
0
27
0
0
0
0
0
0
0
0
21
0
0
0
0
0
0
0
0
3
0
0
0
0
0
0
0
0
6
0
0
0
0
0
0
0
0
61
0
0
0
0
0
0
0
0
25
0
0
0
0
0
0
0
0
39
0
0
0
0
0
0
0
0
30
0
0
0
0
0
0
0
0
11
1
1
1
0
0
0
0
0
36
0 0
0 0
0 0
5 950
E.3.a. Written Language Data (Ranked order)
The META-NET Language Whitepapers 26 also contain Language Tables, which provide a
subjective and qualitative overview of the situation of Language Technologies and Language
Resources for each language.
Again, a detailed analysis of such resources is essential to detect usable resources and the
remaining gaps.
2.7 Maturity in terms of objective evaluations
Objective evaluation can show how "mature" a technology is in terms of performance.
As an example, the following diagram charts the progress of ASR systems over the years, on the
basis of international evaluation campaigns conducted by NIST (the best performance obtained
each year in terms of Word Error Rate (WER)). NIST has conducted multiple evaluations with
various tasks, from voice-activated system with a vocabulary of 1,000 words, voice dictation (5
and 20 K words), radio/TV broadcast news transcriptions (English, Arabic and Chinese
Mandarin) with "high" quality recordings (e.g. no overlap between speakers), telephone
26
http://www.meta-net.eu/meta-share/whitepapers
43
Ch 2. Sustainable LRs
conversations (also in English, Arabic and Mandarin), transcripts of meetings, etc., in variable
conditions.
Timeline for Automatic Speech Recognition since 1987 in terms of NIST evaluation campaigns
27
Today, there are a number of real-world applications in use (transcription of broadcast news,
voice dictation, command and IVR) 28 where the technology achieves sufficient performances. As
it appears in the chart, those performances are similar to those of humans for some tasks (voice
command, voice dictation). But it is also clear that for other more difficult tasks (conversational
speech, or meeting transcriptions), progress is very slow. It should also be stressed that some
tasks, such as voice search, don’t necessitate very high recognition performances. The
performances attained by technology should therefore be compared with the performances
needed by the application.
In the area of speech to speech translation as explored in the TC-STAR project, the results of
evaluation campaigns are as follows: (Joseph comment: The content and the drawings of the
figure are not clear. It may also be said that humans achieve a BLEU score of about 80. And the
results of the “understanding” measure conducted in TC-Star, showing that the level of
understanding by a human of a text translated by a machine is close to the level for the text
translated by a human could also be mentioned.)
27
http://itl.nist.gov/iad/mig/publications/ASRhistory/index.html
Many applications are based on the recognition of audio streams but some (broadcast news) are based on the
“parrot” approach, that involves a well-trained speaker “repeating” (into a speech recognition system adapted to their
voice) the output of a speaker.
28
44
Ch 2. Sustainable LRs
MT Evaluation in the TC-STAR project (Input = Output of automatic speech recognizers, verbatim: transcribed
speech, text= edited transcriptions)
Again, the underlying technologies are now at work in a large number of real-world applications
such as Google-translate.
2.8 Final recommendations
2.14
LRP
2.15
PM
2.16
PM
2.17
PM
2.18
LRP
PM
Draw up a list of the “top 20” technologies – Machine Translation, automatic
speech/speaker recognition, multimodal/multimedia information retrieval/extraction,
summarisation, etc.
Ensure that funding agencies and sponsors put efforts so that a critical mass can be
realised for at least 10 of these technologies, for which crucial resources should be
produced in a framework, publicly and fully funded, that ensure full sharing under the
fairest conditions
Secure an evaluation framework to assess the progress made by such technologies with
respect to state of the art, reported elsewhere
Foster the use of the sustainability model by all resource production players
Ensure that all “best/good practices” are disseminated to the R&D community through
very lightweight methods
45
References
Ch 2. Sustainable LRs
Mariani, Joseph and Francopoulo Gil, (2011). First Public Version of the META-Matrix.
Deliverable D11.1.1. META-NET. www.meta-net.eu/public_documents/t4me/META-NETD11.1.1-Final.pdf
46
Ch.3 Strategies
Chapter 3 - Strategies for Language Resource Mixing,
Sharing, Reusing and Recycling
Stelios Piperidis, Penny Labropoulou
This chapter focuses on reusing and interlinking existing resources and tools as a means to
derive, produce and repurpose new LRs (datasets and relevant processing tools). Together with
the automatic production of LRs (dealt with in Chapter 5), these strategies are ways of
remedying the lack or scarcity of LRs needed for the advancement of LT research and
development.
3.1 Reusing resources
Reuse in this context involves the use of LRs as a component that can be integrated into a new
resource, or extending it in various ways (e.g. linking, translating, etc) to create a new resource;
as seen below, it also involves reusing a LR “model”/specs. “Raw” corpora, for example, can be
partially or fully re-used to create annotated corpora, a tagger can be used as a component of a
corpus management environment, or a morphological lexicon can be combined with other types
of lexicons (encoding subcategorisation frames, sense relations etc.) to create a new, enhanced
lexicon.
New resources have often been produced by re-using existing content. But this has usually been
carried out by the organization that created the original LRs and for their purposes only. Suites
of tools (corpus building components, a POS tagger, parser, annotation tools, statistical data
extraction components etc.) are usually put together to form an integrated corpus management
environment. Similarly, while one part of a large corpus may be morpho-syntactically tagged, a
smaller part may be further parsed to form the basis for a treebank, and the same or yet another
part may be manually or semi-automatically annotated with semantic information. Re-use has
also been extensively applied to the compilation of dictionaries and computational lexicons,
extracting information from corpora (e.g. examples) and other dictionaries (e.g. glosses and
semantic relations).
The classic case for resource re-use by a broad range of organizations and researchers to create
new LRs is WordNet. EuroWordNet was the first project that spurred the creation of new
WordNets for other languages, not only using the basic model but also the resource itself as a
basis for the development of the new LRs. Other languages with different funding sources have
followed.
In addition to its technical merits, the success of WordNet (as mentioned in 2.1) is largely due to
the following
• It is easily accessible and freely available for all types of use.
• A large community, organized as an association, has been created and is active all over
the world supported by a web site dedicated to the constant exchange of information on
wordnets for all languages 1.
• Different methodologies have been devised to develop new language versions depending
on the availability of other resources. The availability of digital monolingual lexical
resources with semantic information encoded in them for certain languages led to the
selection of the merge model, in which the new WordNet is first built independently and
then linked to the most equivalent synset in the Princeton WordNet (i.e. by merging the
1
http://www.globalwordnet.org/
47
Ch.3 Strategies
language-specific wordnets). Then there is the expand model, in which the original
WordNet is expanded with equivalence links from each synset to synsets in the other
languages; this model is favoured for languages in which digital bilingual resources
already exist. In certain cases, the resource producers have even translated the original
WordNet and exploited the original synset relations to create the semantic network of
their own language.
3.2 Interlinking resources
Resources can be interlinked to create new resources.
For instance, interlinking of LRs across languages can be used to produce annotated corpora.
MultiSemCor2, an Italian semantically annotated corpus, has been created by transferring the
annotations of the English original resource (SemCor) to the manually translated version,
following alignment of the two corpora. In fact, using parallel aligned corpora, one of which is
annotated syntactically or semantically and these annotations are then projected on to the other,
is a highly promising method for bootstrapping new LRs with less manual effort 3.
Interlinking different resources within the same language can also help produce larger and
richer LRs. In the case of corpora, interlinking the different annotation layers of corpora can lead
to larger, coherent, multipurpose LRs. The American National Corpus (ANC) 4 has issued an open
invitation for contributions of linguistic annotations on all or part of the ANC or the OANC (the
freely available version of the ANC) and for such derived resources as word lists for free
distribution and use. The OANC has been annotated with various types of linguistic information
(WordNet senses, FrameNet annotations, part of speech tags, Penn Treebank syntactic
annotations etc.). All the annotations are in GrAF format (the ISO standard for standoff
annotation of linguistic data), and can be merged or combined using ANC resources. GrAF
annotations can be loaded into annotation tools such as GATE and UIMA and/or transduced to
other formats using the ANC2Go tool.
As shown by these examples, successful interlinking is facilitated to a large extent by ensuring
interoperability (explored in Ch. 4) which enables the re-usability and mixing of resources and
tools.
3.3 Repurposing resources
Repurposing involves taking a resource originally created for a specific application or domain
and tailoring it to the requirements of another application or domain. A good example would be
taking linguistic/terminological resources and re-engineering them to create ontologies (see e.g.
the NEON project 5). This kind of transfer of resources to new domains and applications, although
it is not easy, seems to be a promising way to reduce production costs and has attracted a lot of
attention recently. As scientific methods and techniques get increasingly used across scientific
areas (or subareas) and problems (e.g. methods used for automatic speech recognition being
carried over to machine translation, but also alignment techniques used in processing both
biological data streams as well as linguistic data) repurposing of resources, e.g. use of
transcribed speech corpus to derive language models for other purposes, is expected to gain
more importance.
2
3
4
5
http://www.multisemcor.fbk.eu
See, for instance, Padó and Lapata. (2009), Bentivogli and Pianta (2005), Hwa et al. (2005), Smith and Eisner (2006).
http://americannnationalcorpus.org
http://www.neon-project.org
48
Ch.3 Strategies
3.4 General considerations on reuse
The examples presented above show some of the ingredients for successful re-use: ease of
access, free availability for research and re-use, clear licensing terms, community building,
exchange of ideas and information across related projects, ease of use, interoperability and
potential for merging, research ideas and experimentation on re-using existing material in
innovative ways.
It is now almost unanimously agreed that the re-use of existing resources can help reduce the
cost and time of LR production. Nevertheless, effective re-use comes with its own kind of pricetag.
The re-use of resources is largely dependent on LR availability, identification and easy access. In
fact, infrastructural issues – such as interoperability, resource sharing, and easy access to LRT –
were recurring messages in all the sessions at the FLaReNet Forums and in other FLaReNet or
FLaReNet-related events, as well as in all three META-NET Vision Groups (Translation and
Localisation6, Media and Information Systems 7, Interactive Systems 8)
A LRT Infrastructure
Many LR groups, initiatives and individuals have been advocating for some time for the creation
of a language resource and technology infrastructure - an open resource infrastructure that
enables the easy sharing of data and tools that can operate seamlessly together. This is now
increasingly recognized as a necessary step for building on each other’s achievements,
integrating resources and technologies, interconnecting and coordinating existing structures
and avoiding the dispersal or inconsistency of various efforts. The META-SHARE infrastructure
is now being implemented by the META-NET (Technologies for the Multilingual European
Information Society) consortium to address this requirement.
Operating this infrastructure efficiently and effectively depends on the will of the LR community,
who must now endorse it, populate it richly with LRs to meet emerging needs and use it.
3.1
LRP
3.2
PM
Make their LRs available, visible and easily accessible through an appropriate
infrastructure (e.g. META-SHARE); participate in building the infrastructure by
providing feedback on relevant actions in this direction
Ensure a stable sustainable infrastructure for LR sharing and exchange; support
continuous development and promotional activities
Legal and licensing issues
Over the last decades we have witnessed a vast growth in digital text and audio-visual content
thanks to the advent of the internet, the large-scale digitization of printed and handwritten
material, and the production of computer and mobile-mediated texts and
multimedia/multimodal data. However, legislation concerning their use is neither clear nor even
consistent across countries. US law includes the fair use act 9, while in a number of European
countries reference is made to the free use of LRs for educational purposes but not for research
purposes, let alone for commercial applications. Moreover, existing legislation, particularly
copyright law, is not uniform over all types of LRs, and there are different regulations regarding
6
7
8
9
http://www.meta-net.eu/vision/reports/VisionGroup-TranslationLocalisation-draft.pdf
http://www.meta-net.eu/vision/reports/VisionGroup-MediaInformationServices-draft.pdf
http://www.meta-net.eu/vision/reports/VisionGroup-InteractiveSystems-draft.pdf
Copyright law of the US: para 107, Title 17.
49
Ch.3 Strategies
the treatment of text as compared with the treatment of audio, images or video, and different
again for software, technology applications and non-creative material in general. In addition, the
rules regarding collective works, databases and works of shared authorship are inconsistent.
Copyright limitations and exceptions need to be harmonised across the EU so that:
a) there are simple and clear rules as to how content that may constitute a LR may be used
b) the restrictions imposed by copyright owners are kept to a minimum and do not hinder
the development of LTs
c) such restrictions are waived in case LRs are deployed for research purposes.
3.3
PM
Harmonize legislation regarding LR use for all types of LRs, and make provisions for
allowing the free use thereof at least for research/non-profit purposes
IPR issues
Clearing the IPR on primary data can create problems for the availability of LRs and should be
considered early in the design and specifications stage. This is often neglected and usually
undertaken at the very end of the production stage, and even afterwards. When more than one
source of primary data is involved, IPR clearance at the final stage can be so time-consuming and
manpower intensive that it is simply never done. Collecting spoken data, for instance, is often
carried out without the prior, legally formulated permission of the informers; and going through
the process of asking informers to sign approvals can at the end of the LR production be more
time-consuming than creating a new LR from scratch.
Nowadays, more and more textual data, especially those collected by individuals for their
personal work, are being harvested without asking prior permission from the owners. Web
crawling is a quick way to obtain large amounts of data for personal research. Inclusion of
personal data (e.g. names of speakers in dialogues) can also open up an LR liable to legal
problems. Derived resources, such as lexicons extracted from corpora, may be more debatable
as to potential copyright law infringement, but they also run the risk of legal prosecution.
LR producers should therefore try to clear the legal rights at kick-off and at the same time make
sure that the material they are collecting can be made available for most types of use, including
modifications and the creation of derivatives. If they can only obtain this for part of the LR, this
should be clearly marked as such and made available to the LR community. When making an LR
available, the original material's IPR should be respected and authorised types of use should be
specified under an appropriate licensing scheme.
3.4
LRP
Clear IPR at the early stages of production; try to ensure that re-use is permitted
3.5
PM
Establish an infrastructure that can help LR producers with legal issues
Collaborative efforts
Due to the growing availability of raw digital data, there is a steady shift of interest towards
annotation, especially at the semantic and pragmatic/discourse levels and across modalities.
Annotated material is extremely valuable for advances in LT; and there should be a focus on
larger volumes and higher quality. As high-level annotation is not yet automated (at least not to
an acceptable quality), other means of producing them could be deployed.
50
Ch.3 Strategies
One solution is to leverage collaborative effort by field experts and trained annotators. Other
methods , such as crowd-sourcing proper via, for instance, the Amazon Mechanical Turk 10, social
tagging and gaming platforms, that rely on eliciting datasets and or content (mainly annotations)
from the general public with little or no training present many advantages; however, there
should be more experimentation, metrics for strengths and drawbacks, and suggestions for
improvements thereof (see Callison-Burch & Drezde 2010), including elaboration of ethical
aspects involved in some of the methods. All collaborative efforts must be co-ordinated to avoid
redundancies and multiply gains by building upon each other's achievements. Communities
should be created and supported by an infrastructure that allows exchange of ideas, comments,
the status of each effort etc., following open invitations for contribution not only of annotated
data (such as those issued by the ANC) but also calls for participation of on-going annotation
projects.
Platforms that enable the storage and parallel processing of shared data are a key success factor.
Standoff annotations or tools that allow existing annotations to be easily stripped off, as well as
tools for merging various annotation levels, will prove valuable in this respect.
3.6
LRP
3.7
PM
Carry out LR production and annotation as collaborative projects; open up existing LRs
for collaborative annotation and the reuse of the annotated results; participate in
communities that carry out similar tasks; evaluate and document their results; develop
new tools and/or adapt existing tools and platforms to the needs of collaborative work
Support collaborative efforts for large LR production and annotation projects through
appropriate funding; support the infrastructure required for collaborative work
Survey/Cataloguing of existing LRs
Before embarking on the production of new LRs from scratch, it is important that LR producers
take existing resources into account. Obvious cases of reuse are extensions/enrichments of
existing LRs, adaptation of tools to new domains and applications and so on. More innovative
ways of reusing LRs have been suggested in recent years, as reusing annotations from other
languages and translations to produce annotations in one's own language as described above.
The roll out of datasets and tools trained on other languages for producing LRs for lessresourced languages should be further investigated. Languages of the same family may show
enough similarities to boost mutual LT advancement by using the same set of datasets and tools.
More efforts should be dedicated to this line of research.
3.8
10
LRP
Check the availability of existing LRs; try to reuse what is available; check what is
available not only for one's own language but also for languages of the same family; try
to see what can be reused in terms of data and tools but also experience; create and
exchange tools that help in reuse/repurposing/interlinking
The Amazon Mechanical Turk builds on the concept of crowd-sourcing (essentially crowd+outsourcing), where, in
general, tasks traditionally performed by an employee or contractor are outsourced via internet to a large group of
people willing to contribute for a low fee. Tasks are normally described as human intelligence tasks, and LR
production, including various annotation tasks, constitute an example.
51
Ch.3 Strategies
3.9
PM
Support reuse projects and research activities in the domain; support activities that
document reuse, formulate guidelines, etc.; fund and promote research activities
around reuse (e.g. dedicated workshops, publications on successful reuse cases, etc.)
Interoperability
Formal and semantic interoperability is a strong desideratum for LR reuse, and standards and
best practices are crucial for ensuring interoperability (see also Chapter 4). The survey carried
out for the FLaReNet deliverable D3.1 11 has demonstrated both the use/usefulness of existing
standards and the adoption of widely used practices as de facto standards. This approach should
be strongly supported. Easy access to standards and best practices is a must. The FLaReNet site
already hosts a repository of standards and best practices 12, including references to resources
that have used them. This initiative should be continued and further enriched. Adoption of
existing standards should be encouraged; areas where there are no standards and/or are under
way, should be identified and the creation of new standards with the widest possible consent of
interested users and field experts should be promoted. To achieve better results, experts from
academia and industry should work together.
3.10
3.11
LRP
Look for standards and best practices that best fit the LRs about to be produced,
already at the early stages of design/specifications; adhere to relevant standards and
best practices; engage in groups that produce standards and provide feedback and
comments; produce LRs that are easily amenable to reuse (e.g. adopt formats that allow
easy reuse)
PM
Support infrastructural activities for collecting and disseminating information on
existing standards and best practices; fund activities for setting up new standards
where they do not exist; fund the development and/or maintenance of tools that
support/enforce/validate standards
Documentation
All documentation relating to a LR should be easily accessible by prospective LR users,
something that is fairly rare today (an exemplary exception is that of resources equipped with
dedicated web sites). ELRA provides documentation for its resources in the ELRA catalogue,
FLaReNet for a few individual resources in its repository 13 but this endeavour needs enrichment.
Any infrastructure that hosts LRs should also host the relevant documentation; the different
types of documentation (specifications, annotation guidelines, user manuals, validation reports,
usage reports etc.) should be clearly identifiable. The META-SHARE infrastructure is taking the
first steps in this direction, integrating many of these documentation dimensions.
3.12
11
12
13
52
LRP
When producing an LR, allocate time and manpower to documentation from the start;
collect and provide documentation (or links to it) when giving access to an LR
See FLaReNet deliverable D3.1 "Report on the scientific, organizational and economic methods and models for
building and maintaining LRs"
http://www.flarenet.eu/?q=FLaReNet_Repository_of_Standards_and_Guidelines
http://www.flarenet.eu/?q=Documentation_about_Individual_Resources
Ch.3 Strategies
3.13
PM
In a LR production project, part of the funding should be allocated to documentation
and dissemination activities; support activities for collecting and storing in appropriate
infrastructures documentation for LRs
Metadata
Metadata descriptions play an important role in making LRs easy to identify. All LRs should
therefore be accompanied with appropriate metadata records. Existing and/or new metadata
schemes should allow sharing information and reducing the time and effort required to produce
these records. LR repositories should be open to metadata harvesting through well-known
protocols.
Formal and semantic interoperability of metadata can be guaranteed through the use of a
common repository, such as the ISO DCR (www.isocat.org). Metadata schemes should take into
consideration the needs of LR users and include information that will help them in identifying,
accessing and using LRs. The META-SHARE and CLARIN metadata schemes have been created
with these considerations in mind, while serving different communities and realms of science
and technology. Appropriate tools for uploading and converting metadata records should be
provided by LR repositories.
3.14
3.15
LRP
PM
Provide appropriate metadata descriptions for all LRs distributed, preferably in one of
the widely-used metadata schemes
Support metadata creation and promotion activities; set guidelines and rules for
metadata description of available LRs and support relevant efforts, including their
normalisation
Evaluation and validation
For LR reuse, it is important to evaluate and validate both content and form to ensure quality.
This means that evaluation and validation practices should be encouraged (see Ch. 5).
Information on this process and its results should be provided in the LR documentation and LR
metadata descriptions.
3.16
PM
Promote evaluation and validation activities of LRs and the dissemination of their
outcomes
Availability
When LRs are made available to the LR community, they should ideally be made available for
most types of use and reuse, including modification, conversion, production of derivatives etc.
Derived resources should also be made available under the same terms as the original resource.
Distribution policies and practices should be revisited in order to fit them better to LT user
requirements. For instance, at the technical level, a corpus should be downloadable so for easier
reuse, instead of merely being queried in a corpus management environment.
At the legal level, restrictions on the use or reuse of LRs should be reduced to a minimum so that
LRs are continuously enriched, and LTs and related services are fully deployed and further
developed. There should be standardized licenses clearly indicating what LR users may do with
53
Ch.3 Strategies
the LRs for all available LRs. ELRA, CLARIN and the META-SHARE network have prepared a set
of licensing templates where reuse is clearly mentioned.
National Funding Agencies should/could support periodic construction of reference resources
for their languages assigning funding dedicated to the clearing of IPR issues so that such
resources can then be made publicly available, distributable, and exploitable.
3.17
3.18
LRP
LRP
PM
Make LRs available for most types of use; make use of standardised licenses where reuse
is clearly stipulated; make LRs accessible in formats that allow reuse
Encourage availability of LRs for multiple types (of re-)use; enforce the creation and
adoption of standardised licenses where reuse is clearly stipulated
Openness
Ideally all LRs should be open to use, reuse, sharing, improvement and deployment in order to
advance research and foster development. LR producers should be encouraged to make their
LRs open at least for research purposes. Entirely publicly-funded LRs should be open or
shareable, but must carry clear attribution and rights-holders information, be properly
documented and be made available with a licence requiring nothing more than attribution.
Funding agencies should be aware of this and funded bodies should agree on legal conditions in
advance.
3.19
LRP
Opt for making LRs open for most uses
3.20
PM
Encourage openness for all LRs; enforce openness for publicly-funded LRs
META-SHARE has aligned itself with the open data movement and prepared a Charter 14, a
Declaration for Language Resource Sharing, which takes into consideration these
recommendations.
References
Bentivogli, L., E. Pianta (2005) "Exploiting parallel texts in the creation of multilingual
semantically annotated resources: the MultiSemCor Corpus". Natural Language Engineering
11(3): 247-261
Callison-Burch, Chris and Dredze, Mark (2010) “Creating Speech and Language Data With
Amazon's Mechanical Turk”. Workshop on Creating Speech and Language Data With
Mechanical Turk at NAACL-HLT, 2010.
Hwa, R., Resnik, P., Weinberg, A., Cabezas, C., and Kolak, O. (2005). Bootstrapping parsers via
syntactic projection across parallel texts. Natural Language Engineering, 11(3).
Padó, S. and Lapata, M. (2009). Cross-lingual annotation projection of semantic roles. Journal of
Artificial Intelligence Research. 36(1): 307-340.
14
54
http://www.meta-net.eu/meta-share/charter
Ch.3 Strategies
Smith, D. A. and Eisner, J. (2006). Quasi-synchronous grammars: Alignment by soft projection of
syntactic dependencies. In Proceedings of the Workshop on Statistical Machine Translation,
pages 23-30, New York
55
Ch.4 Interoperability Framework
Chapter 4 - A Strategic
Interoperability Framework
Action
Plan
for
an
Nicoletta Calzolari, Monica Monachini, Valeria Quochi
Today open, collaborative, shared data is the core of any sound language technology strategy.
And standards are fundamental to exchanging, preserving, maintaining and integrating data and
language resources, and achieving interoperability in any language resource infrastructure.
The older term “reusability” has today evolved into “interoperability” - the ability of information
and communication systems to exchange data and enable the sharing of information and
knowledge. Interoperability was declared one of LT’s major priorities at the first FLaReNet
Forum in Vienna.
An Interoperability Framework can be defined as a dynamic environment of language (and
other) standards and guidelines, where the former can interrelate coherently with one another
and the latter describe how standards work and “speak” to each other. It is also intended to
support the provision of language service interoperability.
Language industry companies need this sort of language strategy because if they cannot
interoperate with their clients and suppliers, they would go out of business: “The lack of
interoperability costs the translation industry a fortune” (report on a recent TAUS survey
(2011b)), referring to the high cost of adjusting file formats and the like, esp. when handing off
tasks in workflows.
4.1
The Current Standard Framework
In the past two decades, there has been increasing awareness of the need to define common
practices and formats for linguistic resources, due to the robustness and industrial-scale
utilisation of certain types of NLP technology.
In the 1990s, several projects laid the foundations for the standardisation of resource
representation and annotation. Among these the Expert Advisory Group on Language
Engineering Standards (EAGLES 1996 1), within which the Corpus Encoding Standard (CES and
XCES, see Ide 1998) was developed, the International Standard for Language Engineering (ISLE 2,
see Calzolari et al. 2002) between EC and NSF, and the SAM project (that led 15 years later to the
SpeechDat family, and developed SAMPA, phonetic alphabet used in the ASR community instead
of IPA). With these projects, Europe was at the forefront in establishing standards for Language
Technology.
Today, standardisation is high on the agenda once again. Consensus has begun to emerge, and in
certain areas stable standards have already been defined. However, for many other areas, work
is still on-going, either because “the emergence of a solid body of web-based standards has
dramatically impacted and re-defined many of our ideas about the ways in which resources will
be stored and accessed over the past several years” (Ide & Romary 2007), or because there are
emerging technologies, such as multimodal systems, that have specific requirements not covered
by existing formats and established practices.
We can therefore observe a continuum of standardisation initiatives at various stages of
consolidation, together with new proposals for standards as various areas of language
technology grow sufficiently mature. While some standards are “official”, i.e. designed and
1
2
http://www.ilc.cnr.it/EAGLES96/home.html
http://www.ilc.cnr.it/EAGLES96/isle/ISLE_Home_Page.htm
56
Ch.4 Interoperability Framework
promoted by standardisation bodies such as ISO 3, W3C 4, (the now defunct) LISA, and ETSI 5,
others are so-called de-facto standards or “best practices”, i.e. widely-used formats and
representation frameworks that have gained community consensus (e.g. WordNet (Fellbaum
1998), PennTreeBank (Marcus et al. 1993), CoNLL (Nivre et al. 2007).
Drawing on a previous report drafted by the CLARIN project (Bel et al. 2009), the original
document has been revised and updated (by FLaReNet together with META-SHARE and ELRA)
with standards relevant for the broader LT community, also addressing those that are typically
used in industry, at different levels of granularity. “The Standards' Landscape Towards an
Interoperability Framework 6” (Bel et al., to appear) thus lists both current standards and ongoing promising standardisation efforts so that the community can monitor and actively
contribute to them. This document is conceived like a “live” document to be adopted and
updated by the community (e.g. in future projects and networks), so as not to restart similar
efforts over and over. It is meant to be a general reference guide for the whole community and
particularly useful for LT initiatives such as the META-SHARE infrastructure, as it provides
concrete indications about standards and best practices that are important for given tasks or
media in LT. These standards are at different stages of development: some are already very well
known and widely used, others more LR-specific standards, especially those developed in the
framework of the ISO Technical Committee 37 devoted to LR management, are in the process of
development or are being revised.
Currently, we can identify a relatively small set of basic standards (defined as “foundational
standards”) that have gained wide consensus and allow basic interoperability and exchange.
These are not necessarily specific to language resources, but provide a minimum basis for
interoperability: e.g. Unicode-UTF8 for character encoding, ISO639 for language codes, W3CXML for textual data, PCM, MP3, ATRAC, for audio, etc.
On top of these come standards that specifically address language resource management and
representation that should also be considered as foundational: ISO 24610-1:2006 - Feature
structure representation, Text Encoding Initiative (TEI), and Lexical Markup Framework (LMF)
for lexical resources (Francopoulo et al. 2006, 2008). They are increasingly recognized as
fundamental for LR interoperability and exchange.
A further set of standards focusing on specific aspects of linguistic and terminological
representation are also currently in place and officially established: TMF (ISO 2002) for
terminology, SynAF (Declerk, 2006) and MAF (Clément and de la Clérgerie, 2005) for
morphological and syntactic annotation.
These standards result from years of work and discussion among groups of experts in various
areas of language technology and are thought to be comprehensive enough to allow for the
representation of most current annotations. Most of them address syntactic interoperability by
providing pivot formats (e.g. LAF/GrAF, Ide and Suderman 2007), while today there is a great
need for semantic interoperability, which is still an actual challenge. Most of the more
linguistically-oriented standards are also periodically revised to make them even more
comprehensive as new technologies appear and new languages are brought into the equation.
They still need considerable effort to promote them and spread awareness to a wider
community.
Standards for terminology management and translation technologies are probably the most
widespread and consolidated of all LR standards, largely due to real market pressure from the
3
4
5
6
http://www.iso.org/iso/standards_development/technical_committees/other_bodies/iso_technical_committee.ht
m?commid=48104
http://www.w3.org/
http://www.etsi.org/
This document also collects input also from the LRE Map, Multilingual Web, the FLaReNet fora, LREC Workshops, ISO
and W3C
57
Ch.4 Interoperability Framework
translation industry. These include: TMF, the overarching foundational standard for all forms of
terminology markup; TMX (Translation Memory eXchange), the vendor-neutral open XML
standard for the exchange of Translation Memory (TM) data created by Computer Aided
Translation (CAT) and localization tools; TBX, the open XML-based standard for exchanging
structured terminological data (that has been approved as an international standard by LISA and
ISO) and now taken over by ETSI (after LISA ceased to exist); XLIFF, an OASIS standard for the
exchange of data for translation. The most recent initiative is the reference architecture OAXAL
(Zydron, 2008), a Open Architecture for XML Authoring and Localization Reference Model, made
up of a number of core standards from W3C, OASIS and LISA (from 2011 ETSI).
Finally, there is a stream of on-going standardisation projects and initiatives focused mainly on
recently mature areas of linguistic analysis and emerging technologies such as semantic
annotation (e.g. SemAF ISO24617 – 1-6) which includes temporal (ISO 24617-1:2009, TimeML)
and space annotation (ISO-Space 7 ISO PWI 24617-4), emotion, i.e. EML (W3C, 2007) and
multimodal annotation, i.e. EMMA (W3C, 2009). The community should monitor these initiatives
closely and actively participate in them. It is recommended that researchers, groups and
companies involved or interested in these areas actively contribute to the definition of such
standards.
In addition to the standards mentioned above, specific communities have established practices
that can be considered as de facto standards, such as WordNet and PennTreeBank. Their
availability, accessibility and openly detailed documentation coupled with the possibility to be
implemented without restrictions are the key to promote their usage, speed-up the development
of modules for their adoption, thus reducing costs and time. A number of tools exist that
facilitate their usage. As these need not change in the near future, it is recommended that
mappers/converters are developed from these best practices/common formats to support other
endorsed/official standards.
To sum up, there are a number of standards that jointly provide a potentially useful framework
ready for adoption.
4.2
Barriers and major problems
There are a number of problems that raise barriers to the broad-based usability of the current
standards framework. Let’s look at some of these problems/barriers, whose analysis calls for a
number of recommendations.
• The key issue that hampers broader usage is not so much a lack of certain standards, but,
especially with respect to LT-specific standards, a lack of (open) tools for using existing
standards easily, and a concomitant lack of tool providers.
•
•
•
7
The lack of (ideally open-source) reference implementations and documentation that
would help others understand clearly what was done and how.
The lack of a developer/user education and culture for using standards. There is a strong
tendency to use idiosyncratic schemes, which cause format incompatibility (even for
minor differences) and prevents annotations to merge or being used together. This in
turn prevents easy reuse of available data.
Within ISO, there are two sets of standards: general interest standards (e.g. country
codes) and special interest standards. General interest standards are free, but the others
are not, and this should be avoided. The problem is that NLP standards are not
considered general interest standards, and therefore have to be paid for, acting as a
brake on wider adoption. There are now attempts – e.g. in ISO TC 37 – to overcome this
https://sites.google.com/site/wikiisospace/
58
Ch.4 Interoperability Framework
•
•
•
situation by allowing direct application of standards free of charge through implemented
databases with free access – e.g. the new ISO 12620 ISOCat (Kemps-Snijders et al. 2009).
In W3C, full documentation of standards is free so it is easy for W3C standards to be
broadly accepted and applied. However, participation in the definition and decisionmaking process for these standards is costly and as a result, there are almost no SMEs
involved. Which means, in turn, that in W3C only the big players can set the rules.
Standards need to be built by consensus which means that standards creation is a slow
process.
As noted in a recent TAUS survey (TAUS/LISA 2011): “the industry lacks an organizing
body or umbrella organization capable of leading the effort and monitoring the
compliance”.
4.1
LRP
Standards must be open
4.2
PM
Need of a body organising, monitoring and coordinating standardisation efforts
4.3
Scenarios for using standards
There are various scenarios that critically involve the need for standards, and they provide
strong motivation for standards adoption and investing in developing those standards that are
still lacking. These include:
− Using the same tools on different resources; using different tools on the same data
In architectures for knowledge mining, or creating new resources, where the same data
have to be used and enriched by (chains of) different tools (Bosma et al. 2009),
8
standards such as common formats become crucial for their success (see KYOTO ).
The use of different tools on the same data is relevant for testing and comparing tools,
and also in collaborative situations to process the same corpus for different purposes.
− Creation of workflows – Web Service Interoperability
Standards will ensure operability in cases where workflows are needed that chain
together tools that were not originally designed to work in a pipeline. Today workflows
can mostly be run with tools that were already designed to work together, or can be
9
enabled by using format converters. Experiences such as TS-STAR and PANACEA (Bel
2010, Toral et al. 2011) show that using a common standardised format will facilitate
integration.
If tools have been built or modified to work directly on common/standard formats,
workflows might be easier to design and quicker to run. Although this is not yet possible,
new tools could naturally move in this direction once the advantages have been
demonstrated.
Workflow management should be generalised to cover both web service interfaces and
local processing.
− Integration/Interlinking of resources, components and services
8
9
www.kyoto-project.eu
http://www.panacea-lr.eu/
59
Ch.4 Interoperability Framework
This has become an important trend for companies providing services. Interoperability
among software components is a major issue today.
When integrating (legacy) resources, there is a need not only for standard formats but
also methodologies and best practices for resource management and updating. In the
10
example of Propbank and PennTreeBank in SemLink , changes in one resource cause
11
problems in linking the resources together, resulting in a lot of extra manual work .
Data lifecycle issues come into play here.
Interoperability is equally critical when integrating new training data sets, and can
facilitate the repurposing of existing data.
− Documentation and metadata
At a different level, there is critical need for standard templates in resource
documentation. This could help users compare available resources, for example.
Consensus on basic sets of Metadata agreed in the community is also of utmost
importance for the easy identification and tracking of resources independently of their
physical location. This is becoming particularly critical in new emerging infrastructures.
There is currently much interest and action in metadata standardisation, not only in
Europe and the USA, but also in Australia.
− Validation of language resources
To achieve certified quality validation of LRs, conformity to an accepted standard is
considered an important criterion (see ELRA Validation Manuals 12).
− Evaluation campaigns: shared tasks
To compare the results of different methods, approaches, or technologies, data must be
encoded and annotated in a common format that different groups can process and use.
Here standards clearly play a fundamental role. In fact, many de-facto standards start
from evaluation campaigns or shared tasks and then spread through the sub-community
13
in question (e.g. CoNLL ).
These initiatives play an important role in encouraging/spreading the use of standards
or best practices and as a means for awareness-raising in a wider community.
− Mashups
Standards are obviously critical for easily integrating data for mashup applications
combining data and functionalities.
− Collaborative creation of resources
Collaborative ways of creating or updating/enhancing language resources represent a
recent trend in the LT community, and standards (should) play an important role in
these as they facilitate the sharing of data and especially annotation and editing tools.
− Preservation
When standards evolve we need to port resources/tools to new emerging standards.
Standards should facilitate this updating process and help avoid mismatches. At the
same time, standards make data preservation easier.
10
11
12
13
60
http://verbs.colorado.edu/semlink/
Martha Palmer 2011 Oral Communication at the SILT Workshop on Interoperability, Brandeis 13-14/4/2011.
http://www.elra.info/Validation-Standards.html
http://www.cnts.ua.ac.be/conll/
Ch.4 Interoperability Framework
4.4
Recommendations and next steps
− Semantic/content interoperability
4.3
Semantic annotation offers improved service-oriented interoperability.
So far we have at most achieved syntactic interoperability, i.e. the ability of systems to
process shared data either directly or using conversion software. Pivot formats such as
GrAF (Ide and Suderman 2007), solve the issue of syntactic interoperability, enabling
formats to merge and convert easily.
Today we desperately need semantic interoperability, the ability of systems to interpret
shared linguistic information in textual, spoken, multimedia content in meaningful,
consistent ways (e.g. with reference to a common set of categories). This is naturally far
more difficult as different natural language coding come into play, as well as different
theoretical linguistic approaches.
LRP
Semantic interoperability is needed
A good methodology of work, already defined in EAGLES, was to define the standard as
the lowest common denominator.
4.4
LRP
Define the standard as the lowest common denominator, at the highest level of
granularity, a basic principle that was already factored into EAGLES
− Linked Data and Open Data
4.5
LRP
Closely monitor the Linked Open Data initiative tightly connected to semantic
interoperability
Interoperability via Linked Data could, for example, be able to link our own objects with
corresponding objects in other fields, and achieve both in-domain and extra-domain
convergence.
− Tools that make it possible to use standards
4.6
LRP
PM
Encourage building tools that enable the use of standards, and step up the availability
of sharable/exchangeable data
− Web services platforms
Web service platforms offer an optimal test case for interoperability and could be used
to showcase critical needs and advantages. So, we should
4.7
LRP
Use the web service model to provide platforms with NLP modules as web services for
various applications
61
Ch.4 Interoperability Framework
Projects such as Language Grid, U-Compare, and PANACEA14 can be understood as an
abstract model for platforms providing services based on LT.
These platforms need both syntactic and semantic interoperability. They can be
evaluated to see if service architectures are the best way to address interoperability
issues.
4.8
LRP
PM
Projects results could be provided as web-services. Cloud-based service architectures
could also be leveraged as enablers for LT development
− Collaborative creation of resources and crowd-sourcing
Using the collaborative paradigm to create language resources could become the fast
lane to standardisation, and at the same time share the cost of resource creation.
Crowd-sourcing with respect to shared resources may be linked to interoperability, as it
requires commonly accepted specifications.
Such a collaborative development of resources would ultimately create a new culture of
joint research.
4.9
LRP
PM
Promote collaborative development of resources also as a help to standardisation
− A Large-scale Collaborative Multilingual Annotation Plan
4.10
LRP
Design a massive multilingual annotation plan, whereby everyone can deposit
annotated data at every possible level of annotation (possibly for parallel/comparable
resources), as a specific type of collaborative initiative for resource annotation
This collaborative approach to creating extremely large amounts of annotated data
would help maximise the visibility, reuse and use of resources annotated according to
common standards and best practices, while at the same time encouraging more
exploration of the diversity of linguistic data. A huge multilingual annotation pool, where
everyone can deposit data annotated at every possible different linguistic level for the
same resources, or for diverse resources, should be defined as a specific type of
collaborative initiative for resource annotation (Calzolari 2010). This could create a
fruitful (community-driven) loop linking the most widely-used annotation schemes to
best practices. This sort of initiative would also be extremely beneficial for METASHARE.
− Evaluation campaigns and validation of resources
14
62
http://www.panacea-lr.eu/
Ch.4 Interoperability Framework
As mentioned in section 4.3, evaluation campaigns help the standards agenda. But the
lack of any European evaluation body to coordinate and supervise common format
definitions remains a genuine problem (see also Chapter 5).
Shared tasks and shared data largely influence common data formats. This makes shared
tasks a key site for interoperability and standards actions for both resources and
components.
4.11
PM
4.12
LRP
Need of a European evaluation and validation body
Create ‘official’ validators (such as http://validator.oaipmh.com/ or the OLAC
validator) to check compliance of LRs with basic linguistic standards
This could help provide validation services for resources to be showcased via METASHARE.
− Interoperability Challenge
The idea of an “Interoperability Challenge” was raised by Nancy Ide and James
Pustejovsky at a SILT Workshop in April 2011 and could become an international
initiative, using the evaluation campaign model, but aimed specifically at organising and
supporting shared tasks to speed up the dissemination of standards and drive forward
interoperability 15.
The NLP community as a whole should be involved in this shift to interoperability by
building a general challenge involving tasks that explicitly require multiple data formats,
annotation schemes, and processing modules, so that participants would be highly
motivated to adapt, adopt, and use standards and common formats and reach an
understanding of the advantages they offer.
4.13
LRP
Set up an “interoperability challenge” as a collective exercise to evaluate (and possibly
measure) interoperability
− Repository of standards and best practices
A small preparatory initiative was started within FLaReNet 16, but this requires a
dedicated effort and must be organised as a collaborative action, so that it acts as a
culture builder. A repository of standards could obviously be linked to a repository of
standards-compliant open/freely available data to maximise its benefits.
4.14
LRP
Collaboratively build and maintain a repository of standards and best practices, linked
to standards-compliant open/freely available data
− Awareness initiatives
15
16
see https://sites.google.com/site/siltforum/files
http://www.flarenet.eu/?q=Standards_and_Best_Practices
63
Ch.4 Interoperability Framework
4.15
PM
Launch awareness programs
Awareness about the existing standards and the motivations behind them is one of the
key factors for enabling their adoption. Educational programs should therefore be
launched to explain, promote and disseminate standards especially to students and
young researchers (e.g. through tutorials at conferences, summer schools, seminars…).
Steps could be taken to include standardisation in regular university curricula. Also,
effective ways to demonstrate the return of investment (ROI) of interoperability must be
sought. Adapting one’s tools and resources to standardised common formats in fact
requires some investments that players may not be willing to make unless the clearly see
advantages.
− Education and training initiatives
4.16
4.17
LRP
PM
LRP
PM
Set up training initiatives to promote and disseminate standards
Standardisation as part of university curricula
− Standards Watch
4.18
LRP
PM
Create a permanent Standards Observatory or Standards Watch
No mechanism is currently available to watch when a topic/area deserves
standardisation. On the European side there is for example a lack of official
standardisation initiatives for such relevant topics as space annotation and the
representation of Lexicon-Ontology relations. These have an economic potential and the
EC is not present enough in standardisation initiatives around these areas.
− Quality Certificate
The Data Seal of Approval 17 is a label for resource quality intended to provide a
certification for data repositories to keep data visible and accessible and to ensure long
term preservation. For example, CLARIN centres are expected to comply with certain
standards. This is linked to the concept of LR preservation and sustainability (see
Chapter 2).
Infrastructures such as META-SHARE could introduce some mechanism of assigning
quality scores to resources and tools and evaluating them for compliance with
standards/best practices, and making these evaluations public. “Penalty” systems could
also be devised.
17
64
http://www.datasealofapproval.org
Ch.4 Interoperability Framework
4.19
LRP
Define and establish a Quality Certificate or quality score for LRs, to be endorsed by the
community
− Multilingual web content-related standards
4.20
LRP
Link-up/collaborate/be in line with ISO, W3C, LISA and other multilingual web contentrelated standards, which in the case of LT can be seen as more basic levels of
representation, to ensure the (potential) integration of LRT into current and future web
content products
Multilinguality should be incorporated into standards, e.g. ISO standards should be
instantiated and generalised for as many languages as possible, although this is not
always the case today.
A recommendation should be made to standardisation bodies to:
4.21
PM
Allow for testing/applying standards on a multilingual basis
4.22
PM
Standardisation initiatives must be conducted at an international level
4.5
Towards an Interoperability Framework
A recurrent request from industrial players at recent meetings such as the META-NET Vision
Groups and the META-Council, is: “give me the standards and give me open data”.
Standards must not only be usable; they must also be used, otherwise they serve no purpose. An
essential step in promoting such use is to make LR standards operational, i.e. coming up with
“Operational Definitions of Interoperability” in the form of concrete proposals.
One step towards this is to ensure that standards are ‘open’. The minimum requirements for
open standards are availability and accessibility for all, detailed documentation and the potential
to be implemented without restrictions. Publicly available standards with public specifications
are vital for easy adoption (Perens, 2010; Krechmer, 2006).
4.23
LRP
Ensure that all standards are fully operational
This is the key recommendation for an interoperability framework.
We need to outline the basic pre-conditions and the essential steps to be taken, some of which
have already been mentioned above.
4.5.1 Technical conditions
− Common metadata
This is a widely recognised pre-condition in all major infrastructure initiatives: ELRA,
LDC, CLARIN, META-SHARE, and of course FLaReNet.
− Explicit semantics of annotation metadata and semantic interoperability
65
Ch.4 Interoperability Framework
It is essential to have an explicit semantics of annotation metadata or data categories.
One mechanism for this is ISOCat, which is currently the only available instrument, even
if there are still many problems with it. Categories can be defined at a persistent web
location that can be referenced from any scheme that uses them.
High-level metadata is not the only set of values that are recorded in ISOCat. So far, the
linguistic categories within ISOCat have been mostly taken from the EAGLES, MULTEXTEast and LIRICS projects (e.g. morpho-syntax, also extended to Semitic, Asian and African
languages), and in the case of terminology from LISA and ISO-12620 value sets. Recently,
ISOCat has been enriched by the CLARIN project in the field of Social Sciences and
Humanities, but these metadata are not precise enough for full-blooded NLP.
This gap between a specific domain and general NLP is currently being addressed by
META-SHARE and it will require the efforts of a broad community of resource
developers/users. To increase convergence towards common data categories, it would
be advisable to create “data category selections” for major standards/best practices. This
would be one step towards semantic interoperability.
Another step would be to get funding agencies to encourage the entering of data
categories and selections in ISOCat, which will only become useful when it is used on a
broad front.
− Tools that help people to use standards
It is extremely important to develop (online) tools that mask the complexities of
standard formats and allow standards to be used easily and support easy
exportation/mapping of data to standards. We recommend the development of
mappers/converters from/to the major standards/best practices/common formats to
the other endorsed/official standards and/or to major proprietary formats.
This is true in particular for infrastructures like META-SHARE where best practices
should also be promoted through tools.
4.5.2 Infrastructural conditions
− A common (virtual) repository for finding the most appropriate standards easily
A joint international effort should organize the indexing of different standards and best
practices, make it easy to find them and keep track of their status, versions and history.
This is critical for an infrastructure such as META-SHARE whose job should be to
recommend standards and best practices for the resources they showcase, especially the
latest examples (the new ones).
− Common documentation templates
Documentation is often inadequate – either too little or too much. This sort of
documentation is essential for a common understanding of standards.
A consensus-driven set of templates for resource documentation should be devised and
disseminated, with actions to facilitate adoption.
− Provide a framework that facilitates testing
We need test scenarios to verify compliance to standards.
− An interoperability framework for/of web services
Making standards operational also means grounding them on an interoperability
framework for/of web services. This means providing standards-compliant linguistic
services, particularly in workflows. In this respect the KYOTO project is both a case study
and a success story.
− “Meta-interoperability” among standards
66
Ch.4 Interoperability Framework
Standards must themselves form a coherent framework, and speak with each other
within an LR-specific ecology.
4.24
LRP
Ensure “meta-interoperability” among standards, that must form a coherent
framework, in a LR-ecology
4.5.3 Social and cultural conditions
The social and/or cultural conditions on interoperability are just as important as the technical
conditions.
− Community involvement
The community as a whole must be involved in standardisation processes. We
recommend that researchers, agencies and companies involved or interested in resource
development/annotation actively help in defining LT standards.
There is a need for an effort to change the old-style “community” mind-set into a new
“social network” mind-set of collaboration to achieve real interoperability. The wider the
participation in such initiatives, the more robust and valid the standards will be.
− Dissemination but not policing
We need to disseminate standardisation, push standards, and invent incentives for using
standards. But people should not feel obliged to conform. Standards must not be seen as
another overhead; people should use them because it is in their interest.
− Interoperability as a valid research area
Interoperability and standardisation issues should become academically acceptable
research areas.
− Link to sustainability
In general, a virtuous circle must be established between standards, use, feedback, and
interoperability.
References
Bel,
N.
et
al.
(2009).
CLARIN
Standardisation
Action
Plan.
CLARIN
http://www.clarin.eu/node/2841
Calzolari, N. et al. (2011). The Standards' Landscape Towards an Interoperability Framework.
FLaReNet Report.
Bel, N. (2010). “Platform for Automatic Normalized Annotation and Cost-Effective Acquisition of
Language Resources for Human Language Technologies: PANACEA”. In Proceedings of the
26th Annual Congress of the Spanish Society for Natural Language Processing (SEPLN),
Valencia.
Bosma, W., et al. (2009). KAF: a generic semantic an-notation format. In: Proceedings of the 5th
International Conference on Generative Approaches to the Lexicon (GL 2009). Pisa.
Calzolari, N., et al. (2002). “Broadening the scope of the EAGLES/ISLE lexical standardization
initiative”. In Proceedings of the 3rd workshop on Asian language resources and international
standardization (COLING '02), vol. 12. pages 1-8, Taipei.
Calzolari N., et al., eds. (2009). Shaping the Future of the Multilingual Digital Europe, 1st
FLaReNet Forum, Vienna.
67
Ch.4 Interoperability Framework
Calzolari, N. (2010). Invited presentation at the COLING 2010 Panel. 23rd International
Conference on Computational Linguistics (COLING 2010). 23-27 August, Beijing, China.
Declerk, T. (2006). “SynAF: Towards a Standard for Syntactic Annotation”. In Proceedings of the
5th International Conference on Language Resources and Evaluation (LREC’06). Pp. 229-232,
Genoa.
EAGLES 1996. http://www.ilc.cnr.it/EAGLES96/home.html
Fellbaum C., ed. (1998). WordNet: An Electronic Lexical Database. Cambridge, MA: MIT Press.
Francopoulo, G. et al. (2006). Lexical markup frame-work (LMF). In Proceedings of the 5th
Interna-tional Conference on Language Resources and Evaluation (LREC'06). pages 233-236,
Genoa.
Francopoulo G. et al. (2009). Multilingual resources for NLP in the lexical markup framework
(LMF). Language Resources and Evaluation. 43(1): 57-70.
Ide, N. (1998). “Corpus Encoding Standard: SGML guidelines for encoding linguistic corpora.” In
Proceedings of the First International Language Resources and Evaluation Conference
(LREC’98). pages 463-470, Granada.
Ide, N and L. Romary (2007). Towards International Standards for Language Resources. In
Dybkjaer, L., Hemsen, H., Minker, W. (eds.), Evaluation of Text and Speech Systems, Springer,
Dordrecht, 263-84.
Ide, N. and K, Suderman. (2007). “GrAF: A graph-based format for linguistic annotations”. In
Proceedings of the Linguistic Annotation Workshop at ACL 2007, pp. 1-8, Prague.
International Organization for Standardisation. 2002. ISO:16642-2002. Terminological Markup
Frame-work. http://www.loria.fr/projets/TMF/
International Organization for Standardization (2008). ISO DIS 24611 Language Resource
Management - Morpho-syntactic Annotation Framework (MAF). ISO/TC 37/SC4/WG 2.
International Organization for Standardization (2008). ISO DIS 24611- (1,2,3,4,5,6) Language
Resource Management - Semantic annotation framework (SemAF). ISO/TC 37/SC4/WG 2.
Kemps-Snijders M., et al (2009). “ISOcat: Remodeling Metadata for Language Resources”.
International Journal of Metadata, Semantics and Ontologies (IJMSO), 4(4): 261-276.
Krechmer K. 2006, Open Standards Requirements. The International Journal of IT Standards and
Standardization Research, 4(1): 43-61.
Lionel C. and Éric de la Clergerie (2005). Maf: a morphosyntactic annotation frame work. In
Proceedings of the 2nd Language and Technology Conference (LTC'05., pages 90-94, Poznan.
Marcus, M. P., B. Santorini, M.A. Marcinkiewicz (1993). Building a Large Annotated Corpus of
English: The Penn Treebank. Computational Linguistics 19 (2): 313-330.
Nivre, J. et al. (2007). The CoNLL 2007 Shared Task on Dependency Parsing. In Proceedings of the
2007 Joint Conference on Empirical Methods in Natural Language Processing and
Computational Natural Language Learning (EMNLP-CoNLL). pages 915-932, Prague.
TAUS (2011a) Annual Plan 2011: Translation Innovation Think Tank Interoperability Watchdog.
http://www.translationautomation.com/images/stories/pdf/taus-annual-plan-2011extended.pdf
TAUS (2011b) Report on a TAUS research about translation interoperability. February 25, 2011.
http://www.translationautomation.com
Toral A., et al. (2011). “Towards a user-friendly web-service architecture for statistical machine
transla-tion in the PANACEA project”. In: M. L. Forcada, H. Depraetere, V. Vandeghinste (eds.)
Proceedings of the 15th EAMT 2011, pages 63-70, Leuven.
W3C (2007) EML: Emotion Incubator Group, W3C Incubator Group Report, 10 July 2007.
W3C (2009) EMMA: Extensible MultiModal Annotation markup language, W3C
Recommendation, 10 February 2009.
68
Ch.4 Interoperability Framework
Zydron A. (2008). OAXAL. What Is It and Why Should I Care? Globalization Insider.
http://www.lisa.org/globalizationinsider/2008.
69
Ch.5 Evaluation & Validation
Chapter 5 - The Evaluation and Validation of Resources
Jan Odijk
This chapter looks at issues of Language Resource (LR) quality in terms of validation and
evaluation, and describes and clarifies recommendations for these activities.
Validation of an LR means checking whether the LR is compliant with its specification and/or
documentation, or, put otherwise, whether "you built the LR right". Validation determines
resource quality (see section 1.6.2). Validation can be formal ("is the form of the LR compatible
with its specification/documentation?") and/or related to content ("is the content of the LR
compatible with its specification/documentation?"). In principle, validation can be applied to
both data and tools/technologies/components. In practice, however, validation has been mainly
applied to data, while evaluation has been mainly applied to tools and technologies.
Evaluation of an LR (broadly construed as including technologies and applications) is checking
whether the LR is suited for a specific task, or, put otherwise, whether "you built the right LR".
Evaluation determines resource adequacy (see section 1.6.3). Though evaluation can in principle
be applied both to data and to technologies and tools, in actual practice evaluation is mostly
applied to technologies and applications only.
5.1
Recommendations for Validation
Validation of an LR during the production process (internal validation) is universally considered
to be essential for ensuring the highest quality of the LR.
Validation of LRs by external parties after the production of the LR (external validation) is
generally (though not universally) considered to be important.
Validation of LRs is possible only if an appropriate specification or documentation of the LR is
available. The quality of LR specifications and documentations tend to vary greatly, and even for
high quality specifications and documentation, the form can vary greatly.
A good approach to achieve actual use of such requirements and to achieve homogeneity in form
is to use templates for LR specifications and documentation. Defining such templates for LRs in
general is perhaps possible, but not sufficiently specific. So templates should be defined for
specifications and documentation of specific LR types (see REC 1.1 and 1.3)
The experiences in the SPEECHDAT family of LR production projects, in which such templates
where defined for certain types of speech resources and in which these templates have actually
been used, proves that such an approach is feasible and yields high quality and highly uniform
specifications and documentation.
Though almost everybody agrees that validation is important, it requires effort and costs. It will
therefore only be done if it is planned and budgeted for in the project plan for the creation of a
resource:
5.1
LRP
Plan and budget internal and external validation in the LR production plan
It is also generally agreed that the effort and costs to carry out validation of LRs must be
minimized.
70
Ch.5 Evaluation & Valdation
5.2
LRP
Minimize the effort and cost of LR validation while maintaining high quality
Even with minimized effort and costs, some believe that external validation is not really
affordable. It is even sometimes thought that you don’t need to validate a resource at all: if a
successful technology can be created on the basis of such a resource, then the resource must be
of a sufficiently high quality, it is claimed. But that is not so clear. In actual practice, all
technology developers have to make adaptations or improvements to resources, or ensure that
errors in resources do not disrupt the technology development, if such resources do not satisfy a
minimum of requirements checked during the validation process. Such effort and costs are
hidden in the research or development costs and they are repeated over and over in every
research or development centre. Validation of resources therefore contributes to minimizing
effort and cost in research and development, and might very well be overall more cost-effective
than no validation.
Interoperability
Several methods for minimizing validation effort and costs are imaginable. Some have been
proposed and should be promoted: An important method is by ensuring interoperability of the
LR with other LRs:
5.3
LRP
5.4
LRP
Ensure formal and semantic interoperability of LRs (see Ch. 4)
With interoperable LRs, large parts of what usually fall under validation (cf. e.g. the ELRA
Validation manuals 1) can be replaced by fully automatic interoperability tests:
Replace (aspects of) validation by fully automatic interoperability tests.
Develop test scenarios for optimally using automatic interoperability tests
If a tool or dataset can be used directly –without any ad-hoc adaptations -- in a pipeline with
other tools and data, this mere fact is an indicator for its quality in certain respects. For example,
if the output of a newly-developed PoS tagger can be used directly as input to an existing
syntactic parser, this is an indicator for the correctness of the PoS-tagger output format and
interpretation (incl. the PoS-tags used). By using a test scenario where a given resource is
combined with a range of other resources, each with its own functionality, multiple aspects of
the quality of the resource are tested. This can be done not only for the resources themselves,
but also for their metadata: one can use tools to check compatibility with metadata harvesting
schemes 2 but one can also use metadata harvesting itself as a test.
To achieve interoperability, several conditions have to be met (see also Ch. 3 and 4), for which
recommendations are given here:
5.5
1
2
LRP
Make the LR visible and accessible via a data and tools infrastructure or exchange
facility
http://www.elra.info/Validation-Standards.html
http://validator.oaipmh.com/
71
Ch.5 Evaluation & Validation
5.6
LRP
5.7
LRP
5.8
LRP
5.9
LRP
5.10
LRP
Use widely accepted de facto or official standards and best practices for data formats
that are supported by the infrastructure and exchange facility in which the LR should
function
Provide metadata for LRs in accordance with widely accepted standards, e.g. Athens
Core 3 in the context of the Component-based Metadata Infrastructure (CMDI) 4
Use formalized metadata elements as much as possible (i.e. metadata elements with a
restricted set of possible values, specific types, etc.) and use the String type for metadata
element values only when it really cannot be avoided
Put all aspects of the documentation of the LR that can be formalized in formalized
metadata elements
Ensure that all data categories used in the metadata and in the data are registered in a
data category registry to ensure semantic interoperability
Automating validation
Several tools, often derived from earlier technology developments, can be used to automate the
detection of errors or likely errors in LRs.
5.11
LRP
Develop new and/or ensure the maximal use of existing tools for automatic or semiautomatic validation of LRs
In the field of spoken LRs, there are tools for fault detection (clipping, noise...), detecting
segmentation errors, of weak annotations, providing confidence measures of speech
transcriptions, etc.
For written resources there are tools to create simple frequency tables, pivot tables etc. for
attributes in the LR and their values, tests of the LRs against lists of permissible values,
aggregation functions and clustering functions that are very useful for detecting real or likely
errors.
We need to validate LRs both formally but also in terms of their content. Content validation is
less easy to automate and therefore it is less easy to minimize effort and cost. This should be a
continuous area of research focus.
5.12
PM
Promote the development of methods for maximally automating content validation
Tools such as the ones mentioned in connection with Rec. 5.11 can be used for formal validation
and also partially for content validation.
3
4
http://www.isocat.org/rest/group/3733
http://www.clarin.eu/cmdi
72
Ch.5 Evaluation & Valdation
These tools are essential for creating large-volume resources that are also of high quality, and
therefore require continuous attention.
5.13
PM
Promote the use of tools that can contribute to ensuring quality in the (automatic)
production of large scale resources
A Collective Effort
Work on validation, its methodologies and supporting tools should be done in the whole
community, and results should be shared to accelerate progress. It is essential that this happens
not only in the academia, but also in the industry:
5.14
PM
Support academic and industry involvement in research on automatic methods for
validation of LRs
Many aspects of LR validation and LR quality require constant adaptation to new developments
(e.g. support for standards, recommendations on metadata issues, recommendations on data
category issues, etc.,). To this end, think tanks should be set up.
5.15
LRP
5.16
LRP
Create a think tank with recognized experts from a broad spectrum of the community
(academia/industry; technologies; and modalities (written/speech/multimodal), etc.)
to assess requirements for LR quality
Enable users to provide comments and votes, both for resources/tools and standards
Past projects, especially the SPEECHDAT family of resource-creation projects, have taught us
that the cooperative creation of specific LRs (e.g. in a consortium where each partner makes an
LR for one language) with clearly specified and agreed requirements and specifications and
resource exchange among partners, is an excellent way to obtain large, high quality LRs for a
wide range of languages as efficiently as possible.
5.17
LRP
Create LRs in collaborative projects where resources are exchanged among project
participants after production
When set up properly, collaborative projects like this one will very often arrive at de facto
standards for data formats that are supported by tools, and are accepted and endorsed by both
academia and industry. Setting up such collaborative projects should be easier nowadays than a
decade and a half ago due to easier interactivity over the internet.
Certain aspects of validation cannot be automated because they require human knowledge.
Some of these ‘Human Intelligence Tasks’ could be carried out by crowd sourcing, which is fairly
easy to organize nowadays. This is used already for resource production, and could equally well
be used for resource validation. The task must be made attractive to users, for example by
73
Ch.5 Evaluation & Validation
presenting it as a game (cf. Google Image Labeler 5, or, in a language context, Phrase Detectives 6).
Or it could be made part of a task that has to be done in the context of security type interactions
such as ReCaptcha 7. A further possibility is to use a service such as the Amazon Mechanical Turk 8.
However, there are ethical and legal issues with such services, which require careful
consideration before making use of them.
LRP
5.18
PM
5.2
Use crowd sourcing techniques to carry out validation tasks, but be aware of the ethical
and legal issues associated with these techniques
Recommendations for Evaluation
There is wide consensus that research on LT technologies has largely been driven by systematic
evaluation during the last two decades. Such evaluation:
• Enables researchers to objectively compare approaches and reproduce experiments
• Helps researchers make issues explicit, validate new ideas, and identify missing science
• Provides an important tool for judging funding efficiency and determining the maturity
of developments for a given application.
It is generally agreed that evaluation should remain an important driving force for this research
in the coming years.
5.19
PM
5.20
PM
Promote evaluation as a driving force for research in LT
Almost every research paper is expected to have an evaluation component, of course, but even
so, evaluation is still a fragmented process. To carry out evaluation systematically, research
should (in part) be organized around evaluation tasks and challenges.
Organize (a significant part of) research around evaluation tasks, challenges and
campaigns
As has become clear in recent years, organizing research in this way has additional benefits:
• It creates focus and mass
• It enables the community to create larger, more representative and better quality
evaluation data and tools
•
•
It helps define, promote and use standards and best practices
Evaluation campaigns have created awareness of the fact that each HLT system has its
own strengths and weaknesses, and that by combining multiple HLT systems using a
voting system, a new one can be created that outperforms each of the input HLT systems.
5
http://images.google.com/imagelabeler/ based on Luis von Ahn’s ESP Game
(http://en.wikipedia.org/wiki/ESP_Game )
6
http://www.phrasedetectives.org
7
8
http://www.google.com/recaptcha
https://www.mturk.com/mturk/welcome
74
Ch.5 Evaluation & Valdation
Such a combined system can then be used to create higher quality annotations than
before.
But many researchers believe that this is not yet enough, and recommend the following:
5.21
LRP
Set up an evaluation management and coordination structure to ensure a viable,
consistent and coherent programme of activities that can successfully scale up and
embrace new communities and technological paradigms
This structure should, among other things,
• Develop a short term (1-year) and midterm (5-year) action plan to take responsibility
for LT evaluation in Europe.
• Propose solutions and actions to include both task- and application-oriented evaluation,
e.g.
o Set-up a general evaluation framework, including both kinds of evaluation,
(comparable to the ISLE Framework for Evaluation in Machine Translation
(FEMTI) approach)
o Set-up an integrated evaluation
• Identify key areas for evaluation and relevant evaluation packages.
• Stimulate work on shared, standard evaluation procedures. Evaluation should include an
assessment of the practical impact of LRT on real NLP applications
• Dedicate special attention to less-resourced languages, and investigate strategies and
funding to include them. In a concerted effort on multiple languages, less-resourced
languages can often “piggy back” on better served languages.
• Dedicate special attention to evaluating technology used for resource production and
acquisition:
o Systematic evaluation of automatic techniques for LR production should be
promoted to assess their strengths and weaknesses and stimulate more research
in these fields
o Appropriate evaluation frameworks for automatic acquisition methods must be
developed to test both current and newly discovered methods.
Since the evaluation of a technology or component under development needs to be carried out
frequently, it is essential that this process is automated as far as possible.
5.22
PM
Promote research into and the development of techniques for fully automatic
evaluation
This very often requires fully formalized evaluation metrics, of which there are many. Some of
these, however, are still very controversial (especially in the area of machine translation) or
considered to be of limited applicability.
5.23
PM
Promote research into and the development of fully formalized evaluation metrics
Many researchers feel that there should be an independent technical infrastructure for
evaluation.
75
Ch.5 Evaluation & Validation
5.24
LRP
Set up a sustainable technical infrastructure providing data, tools and services to carry
out systematic evaluation. This could be a distributed infrastructure involving existing
organizations
This infrastructure should provide
• Ubiquitous remote evaluation
• Evaluation for single components, and also of their contribution to a more global
technology or application.
• Evaluation packages, including tools (e.g. for computing evaluation metrics) and
evaluation data
• Visibility, accessibility and the long term preservation of the components, data and
evaluation packages it hosts. Certification of each participating organization in the
infrastructure for these via commonly agreed upon guidelines and protocols (e.g. via the
Data Seal of Approval 9) is necessary.
For this sort of infrastructure to be successful, the individual components must be interoperable,
as must the components and the data. It would therefore be necessary to
• Define clear interfaces between individual components, e.g., by shared communication
protocols, and
• Ensure agreement on such interfaces within the research community
• Stimulate the development of conversion tools to overcome differences between
interfaces.
This infrastructure could be set up as a web site using web services and work flow systems. It
would be a natural move to use the META-SHARE exchange facility currently being developed in
the META-NET project as a basis for such an evaluation infrastructure.
This kind of infrastructure could bring additional benefits. The practical need for interfaces
between LT technologies, data, and infrastructure components will directly help introduce de
facto interoperability outside the domain of evaluation itself. This infrastructure should
stimulate open access to shareable components, ideally cost-free based on the mutual sharing
and exchange of components.
Sharing components and data has many advantages:
• It avoids duplication,
• It allows research groups to focus on their specialty
• It creates de facto standards and interoperability
• It enables various approaches to be compared easily
• It enables evaluation on both systems and on their component parts.
This will only work if there are clear agreements that will safeguard the interests of commercial
parties.
5.25
9
LRP
Develop clear agreement templates for sharing components and data, especially to
ensure optimal cooperation from commercial parties while safeguarding their interests
http://www.datasealofapproval.org/
76
Ch.5 Evaluation & Valdation
The agreement models used by TAUS for Translation Memories 10, by ECESS in the context of
speech synthesis 11, and by WebForditas 12 for machine translation can provide inspiration.
This kind of management and coordination structure, including a technical infrastructure based
on META-SHARE, will naturally require its own funding, as was pointed out by several
researchers from the community:
5.26
PM
Provide funds for financing the evaluation management and coordination structure and
its associated technical infrastructure. This may require new funding strategies and
instruments both at the European and at national and regional levels
As mentioned in Chapter 2.8, the level of technology performance needed by an application may
vary depending on the application. It is important to make application and technology
developers aware of this fact, and also to better define the required level of performance for a
given application:
5.27
10
11
12
PM
Make technology and application developers aware of the fact that the required
performance level of a technology depends on the application it is used in
Bring technology and application developers together to define the performance levels
required for common applications
http://www.tausdata.org/
http://www.ecess.eu/
http://www.webforditas.hu/?show=textTab&lang=english
77
Ch.6 Automatic LR Development
Chapter 6 - Strategic Directions for Automating LR
Development
Núria Bel, Valeria Quochi
Automatic LR development means automatically compiling resources, for instance by crawling
parallel corpora, achieving grammar induction or the induction of a “polarity” lexicon (listing
items that represent contrasting positions in human discourse such as “this is good/this is bad”),
or by automatically annotating resources so they can be used for new tasks. Examples would be
part of speech tagging for language modeling or annotations for “time” in texts to train a system
to track events.
As mentioned in Ch. 5, the automation of LR development can provide solutions to the high cost
of manually building and annotating new resources, and therefore to the high investments faced
by commercial NLP systems in Europe when localizing their products. At the same time, it can
make advancement in research and technology development possible and quicker. The high
costs for producing language resource data, which in turn are often critical for the development
of good multilingual technology, seems to be the major factor preventing the full deployment of
NLP technologies, especially in geographic areas characterized by a rich variety of official
languages such as Europe. Furthermore, it is equally expensive to adapt both the content and the
systems that manage it to new applications and communication media such as, for instance, SMS
and Tweets.
A clear example of the effects of LR shortage is related to applications for Sentiment
Analysis/Opinion Mining, which have recently expanded massively due to the availability of
content derived from social networks and web reviews in a variety of languages. The aim is to
check whether people have a positive or negative view of some issue, product or related item
and track trends for a given product or service. The preliminary work goes back to the early
1980s (e.g Jaime Carbonell’s PhD thesis 1979), but real commercial interest surged in the 2000s.
Some of the technologies are based on document classification techniques and reach accuracy of
80%. However, it has been proved that employing techniques that use specific language
resources (such as polarity lexica tuned for the specific domains of application) can raise this
performance to 91.33%. (Liu et al. 2005).
The sector of text-analytics for some time was highly affected by this new application to the
point that some companies focused on that area only (e.g., to mention some: Cymfony 1,
Attensity2, Motivequest 3, Sentiment Metrics 4, Lithium 5, QuarkRank 6). Most of the products,
however, mainly concentrate on the processing of opinion expressed in English, and sentiment
analysis resources are still not available across the language spectrum. Once again, there is ongoing research into automatic solutions to create these resources such as information induction
and the repurposing of existing resources. But there seems to be little overall interest in
leveraging this work over the language spread (in fact, in the time span covered by our survey,
only 4 papers describe the application of such techniques and again not for languages other than
English). This is just a case study that supports the recommendation of promoting and
1
2
3
4
5
6
www.cymfony.com
www.attensity.com
www.motivequest.com
www.sentimentmetrics.com
www.lithium.com
www.quarkrank.com
78
Ch.6 Automatic LR Development
supporting research and experimentation on the production of resources so that potentially
killer applications may approach and satisfy market needs.
A survey was carried out in FLaReNet (see D6.1a and b) to assess the state of the art and the
potential for automation of language data resource production, and to express
recommendations. Although the automatic compilation of speech and multimodal resources is
equally important, they were virtually absent from our survey, except in the form of transcribed
speech, which ultimately boils down to written text 7.
6.1
Surveying the state of the art
FLaReNet surveyed the demand for language data to identify whether automatic methods were
being researched and developed to satisfy this demand, and for which applications or purposes
(D6.1a and b).
The first conclusion is that Machine Translation (MT) has been the most demanding area for LRs
in recent years. Though, handcrafted grammars, monolingual and bilingual lexicons have of
course been in continuous demand for commercial systems for some time; the progressive
success of Statistical Machine Translation (SMT) has led to a strong demand firstly for a large
quantity of raw text (i.e. large monolingual and parallel corpora) and then for a growing volume
of annotated data, such as POS tagged texts, identified named entities, syntactically annotated
data, and similar resources.
While some languages are well-provided for with both raw and annotated data (and/or their
concomitant tools), making it easy to launch SMT systems (as demonstrated for example in
Euromatrix 8), there is an evident lack of data for the entire span of languages and semantic
domains, hence, the research interest and recent efforts in trying to automate resource
production as a means to reduce production costs, i.e. tools that can produce such resources
without (or with the least possible) human intervention, the most important cost factor.
SMT is also an example of the community response to data shortage problems. The
unavailability of parallel corpora (texts in different languages which are a translation of the
other for supplying samples to train the SMT system), in particular translation among all of the
European languages, has fostered research into automatic comparable corpora compilation.
Comparable corpora are a collection of texts in different languages, of similar characteristics, but
not translations of one to the other. Current research looks for measures that automatically
measure the degree of comparability. Thus, SMT can be considered an advanced case study that
can help to predict the future needs of other technologies and applications. As in this specific
case of comparable corpora, industry players will probably pay more attention to technologies
for the automatic production of other resources when addressing smaller markets (language
and domain of use), as accurately hand-crafted LRs are simply too expensive.
In order to paint the current picture of automatic production of language resources, which is still
mainly a research issue, FLaReNet systematically surveyed papers on the
production/acquisition of Language Resources accepted at various LT and Computational
Linguistics conferences worldwide between 2006 and 2010. The conferences surveyed have
been selected as representative among those considered most relevant and prestigious in the
domain. These sources include the European Chapter of the Association for Computational
Linguistics (EACL), North American Chapter of the Association for Computational Linguistics
(NAACL), Association for Computational Linguistics (ACL), International Conference on
Computational Linguistics (COLING) and Language Resources and Evaluation Conference
7
8
To a certain extend this is because topics such as automatic speech or video segmentation are normally dealt with in
more physical than symbolic terms.
http://www.euromatrixplus.net/
79
Ch.6 Automatic LR Development
Number of papers
(LREC). A small number of papers in journals 9 were also reviewed, but we only draw on them for
qualitative analysis.
The aim of this survey was to assess whether:
• there is a critical mass of research in automatic LR production that should be supported
and/or transferred to industry
• research on automatic LR development meets the demand for on-going developments in
applications
• it is possible to pinpoint where the (most promising) results are taking place.
The community has also been consulted, as in the FLaReNet Vienna Forum, where there was a
special session on the production of resources. The conclusions of presentations and discussions
were also examined. The results of a poll at the FLaReNet Barcelona Forum were also taken into
account. Finally, we drew on the presentations and discussions at the Workshop on Methods for
the automatic acquisition of Language Resources and their evaluation methods, held as an LREC
workshop, as well as from direct experience of some of the FLaReNet partners with the
PANACEA project 10.
The survey shows that the most frequent (textual/written) areas of research into the production
of resources are the following (see Table 1 and D6.1 for more details):
• Mono- & bilingual lexica/dictionaries, terminology extraction, Multi Word Unit
identification, Named Entity Recognition
• Written corpus acquisition & annotation
• Lexical information: e.g. subcategorizaton frames, selectional preferences, and lexical
features like countability, gender, semantic roles…
• Semantic resources (WordNets, ontologies, polarity and sentiment lexica)
• Grammar induction
• Treebank production
50
45
40
35
30
25
20
15
10
5
0
2006-2007
2008-2009
2010
Written corpus
acquisition &
annotation
Treebank
acquisition
Mono- and bilingual Lexical information
lexica/dictionaries,
acquisition
terminology
extraction, MWEs
identification
Grammar
acquisition
Semantic resources
acquisition
LRT field
Table 1: Overview papers on automatic acquisition of LRs 2006-2010. Distribution per LRT field.
9
Natural Language Engineering (Cambridge University), Computer Speech and Language (Elsevier Publications),
Journal of Language Resources and Evaluation (Springer Publications), Machine Translation (Springer Publications),
Computational Linguistics (MIT Press Journals).
10
80
www.panacea-lr.eu
Ch.6 Automatic LR Development
Although there are certainly other areas where resource automation can be beneficial, the
survey reports on what the community has selected (through publications at renowned
conferences) as being a convincing profile of the field, and which are likely to become promising
areas for the development of robust automatic resource creation technology in the near and
mid-term future.
6.2
Strategic Directions and Recommendations
In this section, drawing from the outcomes of the survey and of related FLaReNet activities, we
establish some recommendations for strategic actions to be put in force either by the community
or by policy makers towards an extensive automation of LR production, which we believe will
lead to beneficial improvements for the commercial value of LTs and thus will enforce
competitiveness of European LT-based enterprises globally.
6.1
LRP
PM
Ensure support and encourage investments in the area of full automation of the full
range of language data production
The list of active and productive research areas in Table 1 cannot be considered a closed list of
areas to support. As mentioned above, spoken data resources, such as richly annotated dialogue
corpora, are largely under-represented in conference papers, but nonetheless are deemed
essential for new research directions in statistical learning approaches to dialogue management,
context-sensitive interpretation, and context-sensitive speech recognition. There might well be
other areas where at present only manual resources are being produced or where automatic
production means are still delivering poor results and thus are not presented/accepted in
conferences. The support to automation must be considered a strategic investment for the whole
area of NLP applications.
6.2
PM
Investments in basic and long term research on methods for automatic production of
language resources should be increased to achieve results in a broader range of
language resources addressed
Table 1 shows written corpora and dictionary/lexicons as the most addressed resources. There
is some evidence of the interest in the development and annotation of corpora and lexicons for
languages other than English, although English still predominates. Thus, it seems that the
emphasis is not necessarily on new technology developments, but on the gradual rollout of
existing technologies to new languages.
The importance given in terms of the number of papers to domain adaptation is also significant
as it could be understood as the community consensus that the development of general purpose
resources (developed once and for all applications) has been shown to not lead to optimal
results.
In quantitative terms, however, it was the topic of “comparable corpus” acquisition that was
driving efforts in 2010, most probably because of the reasons we have previously mentioned.
For the other categories, only lexical acquisition seems to show steady growth, although
semantic resources seem to have recovered, especially referring to the creation of polarity or
subjective lexica for sentiment analysis, one of the most demanding areas as explained above.
In the case of technologies, data in conference papers (in Table 2) suggest that current research
tackling the complete spectrum of approaches to the total automatic production of language
resources is the predominant paradigm, although (and because of the difficulties found) semiautomatic methods are starting to show up.
81
100
90
80
70
60
50
40
30
20
10
0
Ch.6 Automatic LR Development
supervised
semi-supervised
unsupervised
Table 2: Distribution of conference papers per acquisition methodology (*manual here refers only to
collaborative development and repurposing of LRs
The broad success of supervised machine learning techniques has also influenced automatic
resource production. Table 2 shows that most research work uses supervised technologies.
Because in order to train supervised systems the availability of previously annotated data is a
must, and because systems perform better when large quantities of previously annotated data
are used, a vicious circle is created: to produce new data, a sufficient quantity of previously
annotated, high quality data is needed. To create this initial seed data is a costly (usually
manual) exercise that again is an obstacle for doing the exercise for many different languages or
for tuning systems to different domains.
6.3
6.4
6.5
LRP
PM
LRP
LRP
PM
Promote (shared) gold-standard annotation projects for training supervised methods
Promote more efficient uses of available annotated data through a better grasp of
repurposing and merging techniques
Promote research into new methods and techniques for collecting information from
unstructured (non-annotated) data also through inter-disciplinary research with other
scientific domains working on large volumes of data and monitor current innovative
techniques
Because of the cost of manually building gold-standard resources, the automatically produced
ones also are used for training purposes. The final results highly depend on the quality of these
automatically produced resources. Ensuring such quality, evaluation and correction of these
data is also a cost factor and hence it is also noticeable that research has addressed the
development of techniques to automate and reduce human intervention in the evaluation of
automatically created resources. As a general criterion, methods that maximize precision even at
the price of loosing coverage should be promoted because of the need to minimize, in all cases,
human intervention for evaluating and correcting the results of automatically produced
resources. Nevertheless, the problem of error detection also applies to manually produced
82
Ch.6 Automatic LR Development
resources in general thus, methods for automatically determining the quality, or confidence, of
resources and annotations is also a desiderata for language-resource production in general.
6.6
LRP
6.7
LRP
6.8
LRP
Research on evaluation frameworks/methodologies is needed to measure the impact of
resource/tools dependencies and thus the final quality of the end result (see Ch. 5)
Promote research on automatic techniques for error detection, quality and confidence
assessment
Promote technologies that deliver highly accurate, high-confidence results to reduce the
amount of subsequent revision (human or automated)
Because of the cost of getting enough data for the supervised methods previously mentioned,
unsupervised technologies have also attracted the efforts of researchers. However, the results of
unsupervised methods are still not completely satisfactory. Parallel semi-supervised and active
learning 11 techniques are being investigated as a potential solution for building new, highquality resources with a reduced amount of human intervention. Such methods also appear to
have the potential to offset the fact that available data are not large enough for proper training
and used for bootstrapping additional data. Although still at an early stage, they will possibly
emerge as successful research directions in the near future.
6.9
LRP
Invest in methods such as active learning and semi-supervised learning that deliver
promising results while reducing human intervention
Other proposals to reduce costs of the production of gold-standard data refer to the crowdsourcing trend of producing resources by using cheaper, non-expert annotators, as already
mentioned in Chapter 3. However, issues about how to guarantee the quality of such resources
are still open to discussion as severe inconsistencies in data annotation can make the resource
useless for practical robust applications.
Several industry players have chosen crowd-sourcing as a breakthrough to the shortage
problem. It might look as if the industry is either unaware of research into automatic production
of resources, or simply don’t want to invest in this line of research.
However it is of upmost importance that the industry finds ways of cutting the cost of language
resource production because it consists mostly of SMEs with limited investment capacity12. This
limited investment capacity is also likely to be the reason of the limited research carried out by
these players. The set up and promotion of an “industry-friendly” infrastructure for Language
Resource sharing (see Ch. 3) and for research could also be a good instrument for bringing
companies and academic research together on the same tasks and development lines.
6.10
11
12
LRP
Support the emergence of a LR-sharing infrastructure that can lower the cost of R&D
Active learning is a form of supervised machine learning in which the learning algorithm can interactively query the
user (or some other information source) to obtain the desired outputs at new data points. Semi-supervised learning
is a class of machine learning techniques that make use of both labeled and unlabeled data for training - typically a
small amount of labeled data with a large amount of unlabeled data.
ELRA in 2010 published data about LR stakes (interest?): 48% in academia and 37% research and technology
development in industry. Speech related products are again the main consumer of industrial LR.
83
Ch.6 Automatic LR Development
for new applications in new language resource domains
Finally, in order to bring the research community and industry closer, and thus to boost
technological development and commercial competitiveness of European players, it is important
that technologies are evaluated in real world scenarios. To this end new evaluation settings must
be designed and campaigns organised, with public support (at least initially) to convince the
industry of the maturity and utility of certain technologies.
6.11
LRP
6.12
LRP
Carry out quality evaluation in real-world scenarios
Set up evaluation campaigns to boost automatic evaluation of LR production and
promote them across the relevant industries
References
Liu B., M. Hu, and J. Cheng. (2005). “Opinion observer: analyzing and comparing opinions on the
web”. In WWW2005: the 14th international conference on World Wide Web, pp. 342–351, New
York, NY, USA, 2005. ACM Press.
84
Ch.7 An LR Roadmap
Chapter 7 - Towards an LR Roadmap
Joseph Mariani, Khalid Choukri
LTs are developing at a rate proportional to the difficulty of the research challenges and the
complexity of the problems.
NLP research goes back to the very beginning of computer science in the 1950s. A major
paradigm emerged in the mid-1980s opened up by research into Automatic Speech Recognition,
where approaches based on statistical machine learning outperformed those based on crafting
rule sets as expert systems.
Statistical methods paved the way for DARPA-style comparative evaluation campaigns led by the
NBS (what is now the NIST) starting back in 1987. The growing need to gather large quantities
of data to train systems resulted in the creation of the LDC in 1992 on the general assumption
that in NLP “there’s no data like more data.” 1
Since then, this approach has been extended to other areas of language technology - information
retrieval by search engines, Machine Translation, and more generally human-machine
communication (including Computer Vision). It has led to substantial advances and successful
applications, even though it has by no means solved all the problems of automatic language
processing.
Europe has put a similar effort into creating a Language Resource repository, with the launching
of the European Language Resource Association (ELRA) in 1995, which later promoted LR and
evaluation through the LREC conferences that began in 1998. However, despite several shortterm projects, Europe did not create a permanent evaluation agency comparable to the NIST in
the US, although a number of players including ELRA/ELDA played such a role, on a limited scale
and on short term projects.
To develop language technologies, research and industrial teams need access to a volume of data
commensurate with the operational conditions of the application in mind, and equally
appropriate Language Technology evaluation mechanisms including metrics and a methodology.
Hence the importance of Language Resources and Evaluation (LRE). LT evaluation can measure
the performance of the best systems available and compare it to the performances required by
the application (which may not necessarily be 100% in every case).
Also in light of the pioneering actions of the US and the considerable DARPA funding of this
research, the (American) English language has by far the best language data coverage. As a
result, much of the scientific community works on and reports results on English language
phenomena. Technologies and applications grow more and more advanced for English 2, and in
turn produce yet more data and induce the organization of yet new evaluation campaigns,
including the study of metrics themselves – now a new research topic in itself.
The data issue varies considerably for other languages. Some are relatively well covered when
there are programmes that provide the investments needed to produce data and test systems. In
the US, this is the case for geopolitically significant languages (in Iraq, Afghanistan or the
Balkans) or those involved in human emergencies (such as the Haiti earthquake). Other
countries such as France 3, Germany and The Netherlands in Europe have also funded national
1
2
3
B. Mercer, Arden House, 1985 as quoted by F. Jelinek in his keynote talk “Some of my best friends are linguists”,
LREC 2004, 28 May 2004
See for example the large efforts devoted by IBM to produce the Watson knowledge retrieval system which won the
US Jeopardy! Quiz Show in February 2011, while considering that this system works only for English.
Even though US interest has meant that US funded a Mandarin Chinese Broadcast News speech corpus that is 10
times larger than the corresponding corpus in French created under French national programs.
85
Ch.7 An LR Roadmap
programs that accelerate measurable research. But most of the world’s languages do not have
such support.
Some countries see language as a major political issue, either because they want to promote
their language (e.g. Baltic countries) or because they have a constitutional obligation to preserve
the languages spoken by their citizens (e.g. India and South Africa). They prioritize the
development of language technologies to preserve their languages and ensure communication in
them, even if they have limited financial resource to do this extensively. This sort of political
commitment to LT as a support to multilingualism is not yet typical of the European Commission
and the 27 Member States of the European Union.
Companies such as Google or Microsoft play a dominant role in this framework, as they have
access to a huge amount of data in many different languages, devote considerable resources to
LT, have massive computing power and a direct research-to-application pipeline using a new
business model based on so-called “free” services. The fact that a US company like Google is
delivering some of the most comprehensive Language Technology solutions to support
multilingualism in Europe should raise concern among EU officials.
ELRA, FLaReNet and then META-NET conducted a survey on the national and cross-border R&D
effort in language resources and technologies, and existing data, tools, standards and evaluation
methods. This will be available online and continuously updated (the Language Technology
Initiatives Survey 4) with an updated set of Language Matrixes which compare the status of
resources for various languages based on information provided directly by LR producers or
users.
Those language matrices can be used to identify gaps in LT and LR for different languages and
different domains, and explain why there are no applications in these domains. The gaps may
point to a complete lack of LR and LT, or to the insufficient quality of LR and LT, provided there
is a framework for assessing the quality of the LT in question for that language and that
application.
Using this analysis, together with an estimate of the investment needed to cover a given
language task for a given language, we propose a strategy to fill in these gaps in the most
efficient way, at least for EU languages.
Sources
-
7.1
The LRE Map 5
6
The META-NET Language Matrixes
7
The FLaReNet Wiki Survey of National/Transnational LR initiatives
The META-NET National/Transnational LT Programs Survey
The META-NET Summary of detected gaps
8
The META-NET Language White Papers .
Urgent LR requirements
There are two dimensions to identifying LR requirements:
1. Language coverage or diversity:
4
5
6
7
8
http://www.meta-net.eu
http://www.resourcebook.eu/
http://www.meta-net.eu/
http://www.flarenet.eu/?q=Feedback_from_Contact_Points_on_National_Initiatives_in_the_Area_of_Language_Resour
ces
http://www.meta-net.eu/whitepapers
86
Ch.7 An LR Roadmap
This involves identifying the LRs (data, tools and services) needed to develop LT for a given
language, and how to address those needs and reduce, if not eliminate, the current ‘twospeed’ language landscape
2. Topic coverage or topicality:
This involves identifying the innovative types of language data and tools that constitute the
“language resources of the future” for new/strategic applications and sectors, but which are
still lacking or are only at the research/experimental stage.
Language coverage or diversity
To check the availability of LRs for various languages, we explored the data contained in the
META-NET Language Matrixes for monolingual LRs, and in Euromatrix/Euromatrix+ for
bilingual LRs.
The META-NET Language Matrixes derive from the LRE Map 9, which comprises information
provided by the authors of papers submitted to LT conferences. The initial data obtained from
LREC in 2010 is being enriched with material from subsequent conferences- EMNLP’10,
COLING’10, ACL-HLT’11, Interspeech’2011, Oriental Cocosda 2011 - journals (Language
Resources and Evaluation (LRE)) and will be linked to catalogues such as ELRA or LDC. It may
initially appear somewhat biased towards NLP, but data from speech communication
conferences will be added in due course.
To create the LRE Map, a taxonomy of Language Resources was designed by the LREC 2010
Program Committee. LRs were first divided in 4 major categories: Data, Tools, Evaluation and
Meta-Resources (such as guidelines, standards, metadata), and 4 modalities (Written Language,
Spoken Language, Multimodal/Multimedia and Sign Language). 24 LR types were put forward
(see Mariani and Francopoulo 2011) with an option to add other types. The first analysis was
conducted on the 23 official EU languages, on non-EU languages in Europe (such as Norwegian)
and on EU regional languages (such as Catalan and Basque).
The Euromatrix/EuromatrixPlus Matrixes are available on the corresponding project Web sites 10
and concern Machine Translation.
Language coverage in these Matrixes emerged as follows from those quantitative measures:
- Written Language Data (essentially corpus, lexicon, ontology and grammar/language
Models) varies greatly among languages. The most resourced languages are English
(30% of the identified written language data are in English), followed by French and
German, and then Spanish, Italian, Dutch, Portuguese, Swedish and Romanian. Among
the less resourced languages are Estonian, Finnish, Lithuanian, Slovak, Slovene, Maltese
and Irish. Written Language Data also exist for some non-EU European languages and for
several regional European languages.
- Many types of Written Language Tools exist, including taggers/parsers, annotation tools,
named entity recognizers and tokenizers which form the bulk of these tools. The
languages with most coverage are English (25%), French, German, Spanish and Italian.
Many languages are poorly resourced, although many of the tools are languageindependent.
- Spoken Language Data appear less numerous than Written Language Data. They also
mostly cover corpus, grammar/language models and lexicons. The best-resourced
languages are English, French and German, then Spanish and Dutch.
- There are also fewer Spoken Language Tools than Written Language Tools, with a
preponderance of annotation tools. Most of these tools are language-independent.
9
See Calzolari et al. (2010)
10
http://www.euromatrixplus.net/
87
-
-
-
-
-
Ch.7 An LR Roadmap
Annotation tools are especially crucial for producing Multimedia/Multimodal (MM)
Resources. Those tools are mostly language-independent. The most widely covered
languages for MM Data are English (25%) followed by German.
Evaluation Resources (data, tools, methodologies, packages) exist for some languages, but
certainly not all. English is well-served (40%), followed by French, German, Spanish,
Italian and Dutch, which all have or have had LT evaluation activities and programmes.
Finally, there are Meta-Resources, especially standards and metadata for certain
languages, but again, although most of them are language independent, the most
frequently used language is English (30%).
The situation for bilingual resources (corpora, lexicons, etc.) varies considerably among
languages and language pairs, some of them being very well resourced (English (20% of
all resources)), others well resourced (French, German, Spanish, Italian, Portuguese,
Dutch, Greek, Swedish, Polish (from 10% to 4% of all resources), while others again are
considerably under-resourced (Eastern European languages, Baltic languages, Maltese
and Irish (with only 2 identified resources)).
The performance of MT systems as measured by BLEU scores correlates closely to the
availability of existing LRs for corresponding language pairs, as well as to the linguistic
structure of the languages (some of them being less accurately measured by BLEU
scores) and, of course, to the effort devoted to research on those languages.
Topic Coverage or Topicality
To identify the kind of LRs which are needed for new applications, we then analysed the
Language Matrixes and checked new areas suggested by the authors. Some of those new areas
were already covered under different wording in the initial list that was put forward. Others
correspond to the merger of several suggested types. Others were obviously missed when
compiling the taxonomy: these will be added in new versions of the questionnaire to be
circulated in future conferences. Finally, others correspond to “weak signals” indicating that a
new application needs a new technology, and that LRs are being produced to cover that need, at
least for one language. We found the following, among others:
- New Written Language Tools corresponding to new trends in research and applications
of NLP processing at upper levels: knowledge representation tools, semantic role labellers,
semantic network creators, reasoners, sentiment analysis tools. Some aim at facilitating
text analysis such as frequent item set mining, error miners, key phrase extractors, or
rhetorical structure taggers. Toolkits have been developed for experiments on new
algorithmic approaches (CRF Training Toolkit) or for carrying out complete tasks: surface
realizer, text simplification tools, proficiency testing tools, etc.
- There are a few new types of Written Language Data also for upper level processes such
as a word aligned corpus with a multilingual ontology or event semantics.
- Tools for computer aided language learning (CALL) are a new important trend in Spoken
Language Tools.
- Beside annotation tools crucial for producing Multimedia/Multimodal resources, new
kinds of MM tools are emerging such as MM crosslingual information retrieval, associated
with increasing activity in voice-based video search, and tools in machine learning
applications. Talking heads are also of interest, though they are mostly languageindependent.
88
Ch.7 An LR Roadmap
7.2
Findings
Using the previous study of language matrixes, plus a workshop on less-resourced languages
organized as a satellite event at LTC’09 in Poznan (November 2009 11) and a questionnaire
circulated among FLaReNet members and discussed at the FLaReNet Forum in Barcelona
(February 2010), we have identified several findings which we embody as recommendations to
all LT stakeholders - scientific and industrial communities working in this field, educators, the
EC, Member States, regional governments and research agencies.
Language Diversity
In order to properly ‘cover’ a language:
- Visible cooperation between countries and programs would help set an example and
offer Best Practices to less resourced countries. For example, the definition of a
consensual basic set of language resources (or BLARK) has helped produce sufficiently
high quality technologies for a given language, and also the identification of gaps and
roadmaps.
7.1
LRP
-
-
7.2
11
Produce a public Survey on the LT and LR situation worldwide, based on FLaReNet and
META-NET deliverables
Generally speaking, strong political will (rather than lip-service) about the language
dimension and sufficient funds are a sine qua non.
There must also be awareness that LT and LRs are important.
There should be a critical mass of specialists in the language, requiring the training of
young researchers.
There must be a sufficient background:
o A writing system/ transcription code/ agreed orthography
o Language Resources (sufficient in quantity and quality)
o Tools (especially language independent ones such as those based on statistical
training, and ideally open source)
o Metadata, annotation schemes, standards
o Development platforms
o Evaluation facilities (adapted to the language specificities e.g. in the case of the
machine translation of morphologically-rich languages).
The effort should be financed over the long-term, based on a strong foundation, in
agreement with the complexity of language.
Short-term development of a specific product or service for that language (e.g. as a toy),
should be avoided. Demonstrating applications that require a strong foundation should
be encouraged.
LRP
PM
On the basis of the Language Coverage chart, contact officials in national/regional
governments to explain the situation regarding language preservation and
communication through language, the support that LT can bring and the need of LRs to
develop those technologies
See Mariani et al. (2010).
89
Ch.7 An LR Roadmap
7.3
PM
-
7.4
Dialectal variants and sociolinguistics should also be taken into account. This does not
require very much effort, as there are usually commonalities with the corresponding
“main” language.
LRP
-
-
LRP
7.6
LRP
7.7
-
Analyse the relationship within language families and use joint LR methods to develop
LT within those families. Consider the use of pivot languages within those families,
based on the existence of parallel corpora
Adapt LT from the language in which it exists to a similar one in the same language
family, then improve the quality of the LRs for the specific data in that similar language
through bootstrapping
Evaluation must be conducted for all languages, not only for those which attract most of
the research community in its bid for scientific recognition.
LRP
-
Analyse the dialectal variants of languages and include those variants in the production
of LRs
When a majority language also exists, both should be studied together. It would save
time and effort to handle a family of languages altogether. Porting a LT from one
language to a closely related one can be addressed by focusing at specific levels
(phonology, lexicon, syntax) and by producing the corresponding LRs. This is also true
for MT using pivot languages.
Bootstrapping approaches facilitate the coverage of a language.
7.5
-
Start with the EU, extend to the regions and to Europe as a whole, and out to the Rest of
the World (ROW)
Create and promote on-line evaluation tools using Best Practices, guidelines, protocols
and metrics which already exist for certain languages
The related costs could be shared between the corresponding countries or regions, and
international bodies (such as the European Commission (EC)) which could also ensure
proper coordination (see section 7.3).
The keywords must be Interoperability and Sustainability.
Topic coverage
In order to address the question of LRs for innovative applications:
- Information can be extracted from the data entered in the LRE Map
7.8
PM
-
90
Carry out a rolling survey of conference papers and LRs mentioned in papers. Identify
the “weak signals” indicating new trends in LT and LRs internationally
There is a strong trend towards addressing more semantically focused tasks, such as
machine reading, language understanding, knowledge retrieval and human interaction,
requiring models of semantics or pragmatics. Research is now shifting from sentences to
discourse, documents to dialogue, and artificial to natural interaction. And this means
Ch.7 An LR Roadmap
gathering and annotating the data that would facilitate Machine Learning for the design
of working systems.
LRP
7.9
-
PM
-
LRP
PM
-
LRP
7.10
7.11
-
7.13
Web-based applications are particularly interesting due to the availability of online LRs
for training and also reflecting the operational conditions of the application. These LRs
can then be enriched and the quality of the LT improved as a bootstrapping process.
-
7.14
-
Port text data applications to audio data
LRs must be created from real situations for processing emotions in affective computing,
while also taking intentions into account. There is a need for LRs which include both real
data and perceived data.
PM
Promote the development of perceptually annotated corpora
Given the importance of the effort needed to produce LRs, explore the automatic
extraction of LRs from the Web, which means also checking the quality and liability
(rights) of such resources. (see Chapter 6)
LRP
Use automatic content extraction from the Web, but bear legal issues in mind
LRP
Conduct a study and propose a Code of Ethics (regarding privacy and IPR) for
automatic information extraction on the internet
LRP
Develop tools for anonymizing data
There may be a need to draw up a Code of Ethics for the automatic extraction of
information.
PM
7.15
Encourage the development of language based applications on the Web
Applications which used to be specific to written language, such as information retrieval,
question & answer systems, or machine translation, are now becoming available for
spoken language, with a relatively minor drop in performance when compared to
overall performance. They need large amounts of speech data and can be used for
(crosslingual) search on video and TV broadcast data, and for automatic indexing and
sub-titling/dubbing.
LRP
7.12
Promote the collaborative development of semantically/pragmatically/dialogically
annotated corpora
PM
The question of equal access to information for all, regardless of physical disability, is
now a legal matter that will require cross-media technologies, and therefore requires the
91
Ch.7 An LR Roadmap
production of the corresponding LRs. Similarly, the language barrier for accessing
information may appear as a disability, that also requires cross-lingual LRs.
7.16
7.17
-
7.18
PM
Develop applications to answer legal accessibility regulations in Europe. Produce the
corresponding cross-media resources
LRP
View the inability to communicate in a foreign language also as a disability, that should
have a legal answer
PM
Natural interaction other than in telephone applications will ultimately mean removing
the close-talk microphone interface. It is necessary to carry out more research into
audio-acoustics and design and build especially equipped rooms for applications in
meeting transcription, smart homes or robotics.
LRP
-
The NLP community should establish connections with data produced by other
communities, such as medical or neuroscience data.
LRP
7.19
Produce appropriate spoken language resources to study and develop a more natural
approach to voice input
PM
Establish links with other communities in gaining access to better information on the
existence of LRs in their domains, and exchange Best Practices on handling and sharing
resources
Infrastructural matters
-
7.20
-
7.21
Build the proper LR infrastructure
Linguists could help in obtaining high quality LRs, while crowdsourcing and social
networks could help produce large quantities of LRs collaboratively. The ethical aspects
of schemes such as Amazon’s Mechanical Turk should be investigated.
LRP
-
92
LRP
PM
LRP
7.22
7.23
A permanent infrastructure should exist to ensure the interoperability of LR production
and facilitate the distribution processes.
PM
Use crowdsourcing for developing LRs in many languages and of many different kinds
Propose a Code of Ethics for crowdsourcing, in order to guarantee a sufficient salary
and compliance with labour rights (See Chapter 3)
The technical resources of the LR infrastructure should be improved with better human
computer interfaces for searching and sharing LRs. Web services could help in this
process and a common infrastructure would ensure better interoperability.
LRP
Develop and propose (free) tools and more generally Web services (comparable to the
Ch.7 An LR Roadmap
Language Grid), including evaluation protocols and collaborative workbenches in the
LR infrastructure
-
7.24
The LR ecosystem should be promoted more aggressively, with better recognition for LR
producers, on the model of scientific paper authoring.
LRP
Bring together the main LR stakeholders to find a way to attach a Persistent and
Unique Identifier (PUId) (a Persistent, Unique and International Standard LR number
(ISLRN)) to LRs (this may also lead to attaching such unique identifiers to all
researchers, together with an examination of the underlying ethical issues)
LRP
Track the use of LRs in scientific and professional papers
PM
7.25
PM
7.26
LRP
7.27
LRP
LRP
7.28
-
PM
-
7.30
7.31
7.3
Compute a “Language Resource Impact Factor” for each LR
Give greater recognition to successful LRs and their producers (Prizes, Seals of
Recognition, etc. )
There should be more training in production and use of LRs, and LRs should also be used
more widely in education.
LRP
7.29
Attach information related to the initial designer(s) and further contributor(s) to the
LR
PM
Introduce training in the production and use of LR in Computational Linguistics and
Language Technology curricula
There is general agreement that experience and knowledge should be shared more
effectively to avoid reinventing the wheel in the production of new LRs in different
languages or for different tasks.
LRP
LRP
Share all available tools via a LR infrastructure
Continue to organize tutorials on LR production at conferences such as LREC
International Framework
The planned production of necessary LRs will be a huge effort if we consider the number of LTs
multiplied by the number of languages needed to provide solutions to the problem of
multilingualism.
There is unanimous agreement that this effort should be shared by science, industry and
government, using national funding, and regional funding where appropriate, and an EC
contribution for the European Union share, with the support of the Internet community.
93
Ch.7 An LR Roadmap
To address multilingualism more effectively, the experience and best practices gained for one
language should be leveraged to process others, including the share of information about the
needs of LR, sizes and production costs 12, the use of metadata and standards, the reuse of
development methods and of existing tools, such as annotation tools or evaluation protocols and
metrics, the use of automatic transcription, translation and transliteration tools which emerged
from research, and relationships among languages should be exploited to address language
families as a cluster.
The funding mechanisms for a language could be adapted to the size of the effort necessitated by
the lack of resources existing for that language.
7.32
7.33
LRP
Estimate the cost of producing LRs needed to develop an LT for one language
LRP
Establish a joint, coordinated effort using the proper administrative instrument and an
appropriate network infrastructure
PM
PM
7.34
PM
7.35
PM
Share the effort for the production of LRs between international bodies, such as the EC
for the EU, and individual countries/regions, such as Member States in Europe
Adapt the funding amounts and mechanisms to the lack of existing resources for the
corresponding languages
International, National or Regional agencies should open their programs to foreign participants
(and plan for the coverage of the corresponding cost even if they don’t fund those participants).
7.36
PM
Encourage agencies to open their calls to foreign participants
LRP
Leverage LRs built with a similar approach in a common framework as a multilingual
joint training facility
It is useful to use multilingual data as a knowledge source to conduct multilingual joint-training,
and allow for joint knowledge discovery.
7.37
PM
We are now faced with an increasingly complex R&D landscape, due to the sudden upsurge of
worldwide initiatives in LR production/distribution and LT evaluation. It will be vital to
understand the exact topic coverage or agenda of each of these initiatives and build mutual trust.
7.38
LRP
PM
Establish MoU among existing initiatives, starting with the EC efforts
It is important to constantly identify the nature and volume of resources in all countries, to meet
regularly at dedicated forums such as those organized by FLaReNet, and at conferences such as
12
94
For example, it’s worth knowing that the annotation of one hour of speech data may need from 15 hours to 50
hours, depending on the complexity of the annotation process (from orthographic transcription to more complex
labels) and of the targeted quality.
Ch.7 An LR Roadmap
LREC, and establish a permanent international network of stakeholders (researchers, industry
players, laboratories, administrations etc.).
7.39
7.40
7.41
LRP
Ensure the sustainability of the LRE Map, the Language Matrixes and the National
Initiatives Survey Wiki
LRP
Ensure the sustainability of the LREC conference and of the LRE Journal
LRP
Ensure the sustainability of the International FLaReNet International Contact Points,
also through the Survey Wiki
PM
PM
PM
On this basis, it should be feasible to agree among several countries/regions on a common
strategy (contained in a White Paper, or in a series of Memoranda of Understanding (MoU)).
7.42
7.4
LRP
PM
Write a White paper on LT, including LRs, in preparation for FP8. Distribute it within
Europe and abroad
Strategy
To develop LT effectively for the various applications needed to enhance and support
multilingual communication in Europe and worldwide we need to produce and test the quality
of LRs in different languages. This is a huge challenge and needs an appropriate strategy to
ensure an efficient approach to language coverage.
The LTs available for the languages with most advanced technologies, especially (American)
English, necessitated access to available LR of a sufficient size. So the first agenda item is to list
these LRs and identify how much effort was involved in producing them, with an estimate of LT
performance in relation to LR size. There may be open source tools to produce the LRs, but if
not, they should be produced with the aim of being easily portable to many different types of
languages. These LTs must also be evaluated. Here again, there may be guidelines,
methodologies and software to conduct evaluations, which may be reusable for other languages.
If not, they should be produced and shared, with the same aim of being easily portable to many
languages with different specificities.
The availability of a network infrastructure such as META-SHARE will help distribute existing
data, together with the tools for producing new data and for evaluating LTs. It will also help
produce the data itself in a collaborative way and also carry out LT evaluation.
The production of LRs may need the participation of specialists for various processes. For
example, a LR developed for speech translation will need people to transcribe what is said, and
people to translate the resulting transcription into different languages. Similarly, a semantically
annotated speech corpus would possibly need transcription followed by prosodic, syntactic and
semantic labelling.
Those tasks could be achieved by a large population through the new trend in crowdsourcing
over the web. This would help identify appropriate people capable of addressing a specific
language and allow them to work in-country from home. However, the ethical conditions of
crowdsourcing should be carefully checked so that contributors receive appropriate wages and
legal and the tax regulations are observed in each country.
95
Ch.7 An LR Roadmap
The LT evaluation campaigns would also help in the production of LRs. If several systems from
different laboratories process the same data, their results can be used to annotate the data, by
merging them using a mix of automated decision making, and human intervention when results
are different among laboratories or when automation cannot be trusted due to low confidence
scores. This has already been successfully experimented for morphological or syntactic tagging,
and for speech transcription.
The LT evaluation campaigns would also provide insight into the comparative quality of the
tools used, and on the adequacy of LT for a given application that requires a minimum
performance level (this would contribute to an estimate of its Technological Readiness Level
(TRL)). The tools could then be distributed through the same infrastructure, accompanied by
information on their quality. This would create a virtuous circle of LR production through LT
development and LT evaluation to more LR production and LT improvement.
In the European Union, the cost of this effort should be shared between the European
Commission and EU Member States 13, in agreement with the principle of subsidiarity. The
European Commission would have the primary duty of providing the infrastructure necessary to
carry out shared research, including the specification of data exchange standards, the
collaborative infrastructure, communications facilities for building and maintaining links with
all the stakeholders, and an evaluation infrastructure to assess the results and compare different
systems. The Member States, in connection with the Regions, would have the primary duty of
producing the Language Resources needed for their language(s). The development of the core
LT for this can be shared by the various funding agencies.
7.5
Conclusions and perspectives
Our analysis of LRs and LTs shows that although the English language is relatively well covered,
allowing for the availability of the presently operational state-of-the-art applications, many
other languages lack the necessary LRs to develop LTs and applications of the right quality.
There should therefore be a major coordinated effort to ensure that these LRs become available
for all languages.
A Multilingual Europe Technology Alliance offers one model for this, provided that the LRs
needed for LT development in 23 languages can be made available, and that there is an
infrastructure in place to assess LT performances in those different languages on a regular,
systematic basis, and measure their progress and fit to needs. This entire process should be
translated into a ten-year program combining EC programs and Member State national
programs. The model could later be extended to other language clusters, such as non-EU
European languages, European regional languages and any other country’s language(s), using a
cooperative scheme involving financial contributions from national or regional governments
wishing to make their language(s) LT-ready.
References
Calzolari N., C. Soria, R. Del Gratta, S. Goggi, V. Quochi, I. Russo, K. Choukri, J. Mariani, S. Piperidis,
“The LREC 2010 Resource Map”, LREC’2010, Malta, 19-21 May 2010.
Mariani J., K. Choukri, Z. Vetulani, “Report on the Special joint LTC-FLaReNet session « Getting
Less-Resourced Languages On-Board ! », LTC’09 Conference, Poznan”, FLaReNet, February
2010.
13
96
Such a scheme is presently being experimented in the ERIC agreements accompanying the ESFRI programs, as in
CLARIN on Language Resources for the Human and Social Sciences.
Glossary
Glossary of acronyms
ANC
ASR
BLaRK
BNC
CALL
CLARIN
CRF
DARPA
DCR
EC
ELaRK
ELDA
ELRA
EU
FLaReNet
ISO
LDC
LR(s)
LRE
LRE Journal
LRE Map
LREC
LRPs
LRT(s)
LT(s)
META-NET
MM
MT
NBS
NIST
NLP
OANC
OLAC
PMs
POS
R&D
ROW
SLT
SMT
TAUS
TRL
WER
WSD
American National Corpus
Automatic Speech Recognition
Basic Language Resource Kit
British National Corpus
Computer Aided Language Learning
Common Language Resources and Technology Infrastructure
Conditional Random Field
Defense Advanced Research Projects Agency (US)
Data Category Registry
European Commission
Extended Language Resource Kit
Evaluations and Language resources Distribution Agency
European Language Resources Association
European Union
Fostering Language Resources Network
International Standards Organization
Linguistic Data Consortium (USA)
Language Resource(s)
Language Resources and Evaluation
Language Resources and Evaluation Journal
Language Resource and Evaluation Map
Language Resources and Evaluation Conference
Language Resource Producers
Language Resource(s) and Technology(ies)
Language Technology(ies)
Multilingual Europe Technology Alliance
Multimodal/Multimedia
Machine Translation
National Bureau of Standards (USA) (now NIST)
National Institute of Standards and Technology (formerly NBS) (USA)
Natural Language Processing
freely available ANC
Open Language Archives Community
Policy Makers
Part Of Speech (tags and tagger)
Research and Development
Rest Of the World
Spoken Language Translation
Statistical Machine Translation
Translation Automation Users’ Society
Technology Readiness Level
Word Error Rate
Word Sense Disambiguation
97

Similar documents